1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/MemoryBuiltins.h"
21 #include "llvm/Analysis/Loads.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
24 #include "llvm/Analysis/VectorUtils.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/Statepoint.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/KnownBits.h"
42 #include "llvm/Support/MathExtras.h"
47 using namespace llvm::PatternMatch;
49 const unsigned MaxDepth = 6;
51 // Controls the number of uses of the value searched for possible
52 // dominating comparisons.
53 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
54 cl::Hidden, cl::init(20));
56 // This optimization is known to cause performance regressions is some cases,
57 // keep it under a temporary flag for now.
59 DontImproveNonNegativePhiBits("dont-improve-non-negative-phi-bits",
60 cl::Hidden, cl::init(true));
62 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
63 /// returns the element type's bitwidth.
64 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
65 if (unsigned BitWidth = Ty->getScalarSizeInBits())
68 return DL.getPointerTypeSizeInBits(Ty);
72 // Simplifying using an assume can only be done in a particular control-flow
73 // context (the context instruction provides that context). If an assume and
74 // the context instruction are not in the same block then the DT helps in
75 // figuring out if we can use it.
79 const Instruction *CxtI;
80 const DominatorTree *DT;
81 // Unlike the other analyses, this may be a nullptr because not all clients
82 // provide it currently.
83 OptimizationRemarkEmitter *ORE;
85 /// Set of assumptions that should be excluded from further queries.
86 /// This is because of the potential for mutual recursion to cause
87 /// computeKnownBits to repeatedly visit the same assume intrinsic. The
88 /// classic case of this is assume(x = y), which will attempt to determine
89 /// bits in x from bits in y, which will attempt to determine bits in y from
90 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
91 /// isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
92 /// isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so
94 std::array<const Value *, MaxDepth> Excluded;
97 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
98 const DominatorTree *DT, OptimizationRemarkEmitter *ORE = nullptr)
99 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), NumExcluded(0) {}
101 Query(const Query &Q, const Value *NewExcl)
102 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE),
103 NumExcluded(Q.NumExcluded) {
104 Excluded = Q.Excluded;
105 Excluded[NumExcluded++] = NewExcl;
106 assert(NumExcluded <= Excluded.size());
109 bool isExcluded(const Value *Value) const {
110 if (NumExcluded == 0)
112 auto End = Excluded.begin() + NumExcluded;
113 return std::find(Excluded.begin(), End, Value) != End;
116 } // end anonymous namespace
118 // Given the provided Value and, potentially, a context instruction, return
119 // the preferred context instruction (if any).
120 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
121 // If we've been provided with a context instruction, then use that (provided
122 // it has been inserted).
123 if (CxtI && CxtI->getParent())
126 // If the value is really an already-inserted instruction, then use that.
127 CxtI = dyn_cast<Instruction>(V);
128 if (CxtI && CxtI->getParent())
134 static void computeKnownBits(const Value *V, KnownBits &Known,
135 unsigned Depth, const Query &Q);
137 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
138 const DataLayout &DL, unsigned Depth,
139 AssumptionCache *AC, const Instruction *CxtI,
140 const DominatorTree *DT,
141 OptimizationRemarkEmitter *ORE) {
142 ::computeKnownBits(V, Known, Depth,
143 Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
146 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
147 const DataLayout &DL,
148 AssumptionCache *AC, const Instruction *CxtI,
149 const DominatorTree *DT) {
150 assert(LHS->getType() == RHS->getType() &&
151 "LHS and RHS should have the same type");
152 assert(LHS->getType()->isIntOrIntVectorTy() &&
153 "LHS and RHS should be integers");
154 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
155 KnownBits LHSKnown(IT->getBitWidth());
156 KnownBits RHSKnown(IT->getBitWidth());
157 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
158 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
159 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
162 static void ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
163 unsigned Depth, const Query &Q);
165 void llvm::ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
166 const DataLayout &DL, unsigned Depth,
167 AssumptionCache *AC, const Instruction *CxtI,
168 const DominatorTree *DT) {
169 ::ComputeSignBit(V, KnownZero, KnownOne, Depth,
170 Query(DL, AC, safeCxtI(V, CxtI), DT));
173 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
176 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
178 unsigned Depth, AssumptionCache *AC,
179 const Instruction *CxtI,
180 const DominatorTree *DT) {
181 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
182 Query(DL, AC, safeCxtI(V, CxtI), DT));
185 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
187 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
188 AssumptionCache *AC, const Instruction *CxtI,
189 const DominatorTree *DT) {
190 return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
193 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
195 AssumptionCache *AC, const Instruction *CxtI,
196 const DominatorTree *DT) {
197 bool NonNegative, Negative;
198 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
202 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
203 AssumptionCache *AC, const Instruction *CxtI,
204 const DominatorTree *DT) {
205 if (auto *CI = dyn_cast<ConstantInt>(V))
206 return CI->getValue().isStrictlyPositive();
208 // TODO: We'd doing two recursive queries here. We should factor this such
209 // that only a single query is needed.
210 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
211 isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
214 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
215 AssumptionCache *AC, const Instruction *CxtI,
216 const DominatorTree *DT) {
217 bool NonNegative, Negative;
218 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
222 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
224 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
225 const DataLayout &DL,
226 AssumptionCache *AC, const Instruction *CxtI,
227 const DominatorTree *DT) {
228 return ::isKnownNonEqual(V1, V2, Query(DL, AC,
229 safeCxtI(V1, safeCxtI(V2, CxtI)),
233 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
236 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
237 const DataLayout &DL,
238 unsigned Depth, AssumptionCache *AC,
239 const Instruction *CxtI, const DominatorTree *DT) {
240 return ::MaskedValueIsZero(V, Mask, Depth,
241 Query(DL, AC, safeCxtI(V, CxtI), DT));
244 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
247 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
248 unsigned Depth, AssumptionCache *AC,
249 const Instruction *CxtI,
250 const DominatorTree *DT) {
251 return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
254 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
256 KnownBits &KnownOut, KnownBits &Known2,
257 unsigned Depth, const Query &Q) {
258 unsigned BitWidth = KnownOut.getBitWidth();
260 // If an initial sequence of bits in the result is not needed, the
261 // corresponding bits in the operands are not needed.
262 KnownBits LHSKnown(BitWidth);
263 computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
264 computeKnownBits(Op1, Known2, Depth + 1, Q);
266 // Carry in a 1 for a subtract, rather than a 0.
267 uint64_t CarryIn = 0;
269 // Sum = LHS + ~RHS + 1
270 std::swap(Known2.Zero, Known2.One);
274 APInt PossibleSumZero = ~LHSKnown.Zero + ~Known2.Zero + CarryIn;
275 APInt PossibleSumOne = LHSKnown.One + Known2.One + CarryIn;
277 // Compute known bits of the carry.
278 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnown.Zero ^ Known2.Zero);
279 APInt CarryKnownOne = PossibleSumOne ^ LHSKnown.One ^ Known2.One;
281 // Compute set of known bits (where all three relevant bits are known).
282 APInt LHSKnownUnion = LHSKnown.Zero | LHSKnown.One;
283 APInt RHSKnownUnion = Known2.Zero | Known2.One;
284 APInt CarryKnownUnion = CarryKnownZero | CarryKnownOne;
285 APInt Known = LHSKnownUnion & RHSKnownUnion & CarryKnownUnion;
287 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
288 "known bits of sum differ");
290 // Compute known bits of the result.
291 KnownOut.Zero = ~PossibleSumOne & Known;
292 KnownOut.One = PossibleSumOne & Known;
294 // Are we still trying to solve for the sign bit?
295 if (!Known.isSignBitSet()) {
297 // Adding two non-negative numbers, or subtracting a negative number from
298 // a non-negative one, can't wrap into negative.
299 if (LHSKnown.isNonNegative() && Known2.isNonNegative())
300 KnownOut.makeNonNegative();
301 // Adding two negative numbers, or subtracting a non-negative number from
302 // a negative one, can't wrap into non-negative.
303 else if (LHSKnown.isNegative() && Known2.isNegative())
304 KnownOut.makeNegative();
309 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
310 KnownBits &Known, KnownBits &Known2,
311 unsigned Depth, const Query &Q) {
312 unsigned BitWidth = Known.getBitWidth();
313 computeKnownBits(Op1, Known, Depth + 1, Q);
314 computeKnownBits(Op0, Known2, Depth + 1, Q);
316 bool isKnownNegative = false;
317 bool isKnownNonNegative = false;
318 // If the multiplication is known not to overflow, compute the sign bit.
321 // The product of a number with itself is non-negative.
322 isKnownNonNegative = true;
324 bool isKnownNonNegativeOp1 = Known.isNonNegative();
325 bool isKnownNonNegativeOp0 = Known2.isNonNegative();
326 bool isKnownNegativeOp1 = Known.isNegative();
327 bool isKnownNegativeOp0 = Known2.isNegative();
328 // The product of two numbers with the same sign is non-negative.
329 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
330 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
331 // The product of a negative number and a non-negative number is either
333 if (!isKnownNonNegative)
334 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
335 isKnownNonZero(Op0, Depth, Q)) ||
336 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
337 isKnownNonZero(Op1, Depth, Q));
341 // If low bits are zero in either operand, output low known-0 bits.
342 // Also compute a conservative estimate for high known-0 bits.
343 // More trickiness is possible, but this is sufficient for the
344 // interesting case of alignment computation.
345 unsigned TrailZ = Known.Zero.countTrailingOnes() +
346 Known2.Zero.countTrailingOnes();
347 unsigned LeadZ = std::max(Known.Zero.countLeadingOnes() +
348 Known2.Zero.countLeadingOnes(),
349 BitWidth) - BitWidth;
351 TrailZ = std::min(TrailZ, BitWidth);
352 LeadZ = std::min(LeadZ, BitWidth);
354 Known.Zero.setLowBits(TrailZ);
355 Known.Zero.setHighBits(LeadZ);
357 // Only make use of no-wrap flags if we failed to compute the sign bit
358 // directly. This matters if the multiplication always overflows, in
359 // which case we prefer to follow the result of the direct computation,
360 // though as the program is invoking undefined behaviour we can choose
361 // whatever we like here.
362 if (isKnownNonNegative && !Known.isNegative())
363 Known.makeNonNegative();
364 else if (isKnownNegative && !Known.isNonNegative())
365 Known.makeNegative();
368 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
370 unsigned BitWidth = Known.getBitWidth();
371 unsigned NumRanges = Ranges.getNumOperands() / 2;
372 assert(NumRanges >= 1);
374 Known.Zero.setAllBits();
375 Known.One.setAllBits();
377 for (unsigned i = 0; i < NumRanges; ++i) {
379 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
381 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
382 ConstantRange Range(Lower->getValue(), Upper->getValue());
384 // The first CommonPrefixBits of all values in Range are equal.
385 unsigned CommonPrefixBits =
386 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
388 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
389 Known.One &= Range.getUnsignedMax() & Mask;
390 Known.Zero &= ~Range.getUnsignedMax() & Mask;
394 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
395 SmallVector<const Value *, 16> WorkSet(1, I);
396 SmallPtrSet<const Value *, 32> Visited;
397 SmallPtrSet<const Value *, 16> EphValues;
399 // The instruction defining an assumption's condition itself is always
400 // considered ephemeral to that assumption (even if it has other
401 // non-ephemeral users). See r246696's test case for an example.
402 if (is_contained(I->operands(), E))
405 while (!WorkSet.empty()) {
406 const Value *V = WorkSet.pop_back_val();
407 if (!Visited.insert(V).second)
410 // If all uses of this value are ephemeral, then so is this value.
411 if (all_of(V->users(), [&](const User *U) { return EphValues.count(U); })) {
416 if (const User *U = dyn_cast<User>(V))
417 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
419 if (isSafeToSpeculativelyExecute(*J))
420 WorkSet.push_back(*J);
428 // Is this an intrinsic that cannot be speculated but also cannot trap?
429 static bool isAssumeLikeIntrinsic(const Instruction *I) {
430 if (const CallInst *CI = dyn_cast<CallInst>(I))
431 if (Function *F = CI->getCalledFunction())
432 switch (F->getIntrinsicID()) {
434 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
435 case Intrinsic::assume:
436 case Intrinsic::dbg_declare:
437 case Intrinsic::dbg_value:
438 case Intrinsic::invariant_start:
439 case Intrinsic::invariant_end:
440 case Intrinsic::lifetime_start:
441 case Intrinsic::lifetime_end:
442 case Intrinsic::objectsize:
443 case Intrinsic::ptr_annotation:
444 case Intrinsic::var_annotation:
451 bool llvm::isValidAssumeForContext(const Instruction *Inv,
452 const Instruction *CxtI,
453 const DominatorTree *DT) {
455 // There are two restrictions on the use of an assume:
456 // 1. The assume must dominate the context (or the control flow must
457 // reach the assume whenever it reaches the context).
458 // 2. The context must not be in the assume's set of ephemeral values
459 // (otherwise we will use the assume to prove that the condition
460 // feeding the assume is trivially true, thus causing the removal of
464 if (DT->dominates(Inv, CxtI))
466 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
467 // We don't have a DT, but this trivially dominates.
471 // With or without a DT, the only remaining case we will check is if the
472 // instructions are in the same BB. Give up if that is not the case.
473 if (Inv->getParent() != CxtI->getParent())
476 // If we have a dom tree, then we now know that the assume doens't dominate
477 // the other instruction. If we don't have a dom tree then we can check if
478 // the assume is first in the BB.
480 // Search forward from the assume until we reach the context (or the end
481 // of the block); the common case is that the assume will come first.
482 for (auto I = std::next(BasicBlock::const_iterator(Inv)),
483 IE = Inv->getParent()->end(); I != IE; ++I)
488 // The context comes first, but they're both in the same block. Make sure
489 // there is nothing in between that might interrupt the control flow.
490 for (BasicBlock::const_iterator I =
491 std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
493 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
496 return !isEphemeralValueOf(Inv, CxtI);
499 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
500 unsigned Depth, const Query &Q) {
501 // Use of assumptions is context-sensitive. If we don't have a context, we
503 if (!Q.AC || !Q.CxtI)
506 unsigned BitWidth = Known.getBitWidth();
508 // Note that the patterns below need to be kept in sync with the code
509 // in AssumptionCache::updateAffectedValues.
511 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
514 CallInst *I = cast<CallInst>(AssumeVH);
515 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
516 "Got assumption for the wrong function!");
520 // Warning: This loop can end up being somewhat performance sensetive.
521 // We're running this loop for once for each value queried resulting in a
522 // runtime of ~O(#assumes * #values).
524 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
525 "must be an assume intrinsic");
527 Value *Arg = I->getArgOperand(0);
529 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
530 assert(BitWidth == 1 && "assume operand is not i1?");
534 if (match(Arg, m_Not(m_Specific(V))) &&
535 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
536 assert(BitWidth == 1 && "assume operand is not i1?");
541 // The remaining tests are all recursive, so bail out if we hit the limit.
542 if (Depth == MaxDepth)
546 auto m_V = m_CombineOr(m_Specific(V),
547 m_CombineOr(m_PtrToInt(m_Specific(V)),
548 m_BitCast(m_Specific(V))));
550 CmpInst::Predicate Pred;
553 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
554 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
555 KnownBits RHSKnown(BitWidth);
556 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
557 Known.Zero |= RHSKnown.Zero;
558 Known.One |= RHSKnown.One;
560 } else if (match(Arg,
561 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
562 Pred == ICmpInst::ICMP_EQ &&
563 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
564 KnownBits RHSKnown(BitWidth);
565 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
566 KnownBits MaskKnown(BitWidth);
567 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
569 // For those bits in the mask that are known to be one, we can propagate
570 // known bits from the RHS to V.
571 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
572 Known.One |= RHSKnown.One & MaskKnown.One;
573 // assume(~(v & b) = a)
574 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
576 Pred == ICmpInst::ICMP_EQ &&
577 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
578 KnownBits RHSKnown(BitWidth);
579 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
580 KnownBits MaskKnown(BitWidth);
581 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
583 // For those bits in the mask that are known to be one, we can propagate
584 // inverted known bits from the RHS to V.
585 Known.Zero |= RHSKnown.One & MaskKnown.One;
586 Known.One |= RHSKnown.Zero & MaskKnown.One;
588 } else if (match(Arg,
589 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
590 Pred == ICmpInst::ICMP_EQ &&
591 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
592 KnownBits RHSKnown(BitWidth);
593 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
594 KnownBits BKnown(BitWidth);
595 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
597 // For those bits in B that are known to be zero, we can propagate known
598 // bits from the RHS to V.
599 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
600 Known.One |= RHSKnown.One & BKnown.Zero;
601 // assume(~(v | b) = a)
602 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
604 Pred == ICmpInst::ICMP_EQ &&
605 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
606 KnownBits RHSKnown(BitWidth);
607 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
608 KnownBits BKnown(BitWidth);
609 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
611 // For those bits in B that are known to be zero, we can propagate
612 // inverted known bits from the RHS to V.
613 Known.Zero |= RHSKnown.One & BKnown.Zero;
614 Known.One |= RHSKnown.Zero & BKnown.Zero;
616 } else if (match(Arg,
617 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
618 Pred == ICmpInst::ICMP_EQ &&
619 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
620 KnownBits RHSKnown(BitWidth);
621 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
622 KnownBits BKnown(BitWidth);
623 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
625 // For those bits in B that are known to be zero, we can propagate known
626 // bits from the RHS to V. For those bits in B that are known to be one,
627 // we can propagate inverted known bits from the RHS to V.
628 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
629 Known.One |= RHSKnown.One & BKnown.Zero;
630 Known.Zero |= RHSKnown.One & BKnown.One;
631 Known.One |= RHSKnown.Zero & BKnown.One;
632 // assume(~(v ^ b) = a)
633 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
635 Pred == ICmpInst::ICMP_EQ &&
636 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
637 KnownBits RHSKnown(BitWidth);
638 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
639 KnownBits BKnown(BitWidth);
640 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
642 // For those bits in B that are known to be zero, we can propagate
643 // inverted known bits from the RHS to V. For those bits in B that are
644 // known to be one, we can propagate known bits from the RHS to V.
645 Known.Zero |= RHSKnown.One & BKnown.Zero;
646 Known.One |= RHSKnown.Zero & BKnown.Zero;
647 Known.Zero |= RHSKnown.Zero & BKnown.One;
648 Known.One |= RHSKnown.One & BKnown.One;
649 // assume(v << c = a)
650 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
652 Pred == ICmpInst::ICMP_EQ &&
653 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
654 KnownBits RHSKnown(BitWidth);
655 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
656 // For those bits in RHS that are known, we can propagate them to known
657 // bits in V shifted to the right by C.
658 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
659 Known.Zero |= RHSKnown.Zero;
660 RHSKnown.One.lshrInPlace(C->getZExtValue());
661 Known.One |= RHSKnown.One;
662 // assume(~(v << c) = a)
663 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
665 Pred == ICmpInst::ICMP_EQ &&
666 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
667 KnownBits RHSKnown(BitWidth);
668 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
669 // For those bits in RHS that are known, we can propagate them inverted
670 // to known bits in V shifted to the right by C.
671 RHSKnown.One.lshrInPlace(C->getZExtValue());
672 Known.Zero |= RHSKnown.One;
673 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
674 Known.One |= RHSKnown.Zero;
675 // assume(v >> c = a)
676 } else if (match(Arg,
677 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
678 m_AShr(m_V, m_ConstantInt(C))),
680 Pred == ICmpInst::ICMP_EQ &&
681 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
682 KnownBits RHSKnown(BitWidth);
683 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
684 // For those bits in RHS that are known, we can propagate them to known
685 // bits in V shifted to the right by C.
686 Known.Zero |= RHSKnown.Zero << C->getZExtValue();
687 Known.One |= RHSKnown.One << C->getZExtValue();
688 // assume(~(v >> c) = a)
689 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
690 m_LShr(m_V, m_ConstantInt(C)),
691 m_AShr(m_V, m_ConstantInt(C)))),
693 Pred == ICmpInst::ICMP_EQ &&
694 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
695 KnownBits RHSKnown(BitWidth);
696 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
697 // For those bits in RHS that are known, we can propagate them inverted
698 // to known bits in V shifted to the right by C.
699 Known.Zero |= RHSKnown.One << C->getZExtValue();
700 Known.One |= RHSKnown.Zero << C->getZExtValue();
701 // assume(v >=_s c) where c is non-negative
702 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
703 Pred == ICmpInst::ICMP_SGE &&
704 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
705 KnownBits RHSKnown(BitWidth);
706 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
708 if (RHSKnown.isNonNegative()) {
709 // We know that the sign bit is zero.
710 Known.makeNonNegative();
712 // assume(v >_s c) where c is at least -1.
713 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
714 Pred == ICmpInst::ICMP_SGT &&
715 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
716 KnownBits RHSKnown(BitWidth);
717 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
719 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
720 // We know that the sign bit is zero.
721 Known.makeNonNegative();
723 // assume(v <=_s c) where c is negative
724 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
725 Pred == ICmpInst::ICMP_SLE &&
726 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
727 KnownBits RHSKnown(BitWidth);
728 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
730 if (RHSKnown.isNegative()) {
731 // We know that the sign bit is one.
732 Known.makeNegative();
734 // assume(v <_s c) where c is non-positive
735 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
736 Pred == ICmpInst::ICMP_SLT &&
737 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
738 KnownBits RHSKnown(BitWidth);
739 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
741 if (RHSKnown.isZero() || RHSKnown.isNegative()) {
742 // We know that the sign bit is one.
743 Known.makeNegative();
746 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
747 Pred == ICmpInst::ICMP_ULE &&
748 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
749 KnownBits RHSKnown(BitWidth);
750 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
752 // Whatever high bits in c are zero are known to be zero.
753 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes());
755 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
756 Pred == ICmpInst::ICMP_ULT &&
757 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
758 KnownBits RHSKnown(BitWidth);
759 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
761 // Whatever high bits in c are zero are known to be zero (if c is a power
762 // of 2, then one more).
763 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
764 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes()+1);
766 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes());
770 // If assumptions conflict with each other or previous known bits, then we
771 // have a logical fallacy. It's possible that the assumption is not reachable,
772 // so this isn't a real bug. On the other hand, the program may have undefined
773 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
774 // clear out the known bits, try to warn the user, and hope for the best.
775 if (Known.Zero.intersects(Known.One)) {
779 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
780 OptimizationRemarkAnalysis ORA("value-tracking", "BadAssumption", CxtI);
781 Q.ORE->emit(ORA << "Detected conflicting code assumptions. Program may "
782 "have undefined behavior, or compiler may have "
788 // Compute known bits from a shift operator, including those with a
789 // non-constant shift amount. Known is the outputs of this function. Known2 is a
790 // pre-allocated temporary with the/ same bit width as Known. KZF and KOF are
791 // operator-specific functors that, given the known-zero or known-one bits
792 // respectively, and a shift amount, compute the implied known-zero or known-one
793 // bits of the shift operator's result respectively for that shift amount. The
794 // results from calling KZF and KOF are conservatively combined for all
795 // permitted shift amounts.
796 static void computeKnownBitsFromShiftOperator(
797 const Operator *I, KnownBits &Known, KnownBits &Known2,
798 unsigned Depth, const Query &Q,
799 function_ref<APInt(const APInt &, unsigned)> KZF,
800 function_ref<APInt(const APInt &, unsigned)> KOF) {
801 unsigned BitWidth = Known.getBitWidth();
803 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
804 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
806 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
807 Known.Zero = KZF(Known.Zero, ShiftAmt);
808 Known.One = KOF(Known.One, ShiftAmt);
809 // If there is conflict between Known.Zero and Known.One, this must be an
810 // overflowing left shift, so the shift result is undefined. Clear Known
811 // bits so that other code could propagate this undef.
812 if ((Known.Zero & Known.One) != 0)
818 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
820 // If the shift amount could be greater than or equal to the bit-width of the LHS, the
821 // value could be undef, so we don't know anything about it.
822 if ((~Known.Zero).uge(BitWidth)) {
827 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
828 // BitWidth > 64 and any upper bits are known, we'll end up returning the
829 // limit value (which implies all bits are known).
830 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
831 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
833 // It would be more-clearly correct to use the two temporaries for this
834 // calculation. Reusing the APInts here to prevent unnecessary allocations.
837 // If we know the shifter operand is nonzero, we can sometimes infer more
838 // known bits. However this is expensive to compute, so be lazy about it and
839 // only compute it when absolutely necessary.
840 Optional<bool> ShifterOperandIsNonZero;
842 // Early exit if we can't constrain any well-defined shift amount.
843 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
844 ShifterOperandIsNonZero =
845 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
846 if (!*ShifterOperandIsNonZero)
850 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
852 Known.Zero.setAllBits();
853 Known.One.setAllBits();
854 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
855 // Combine the shifted known input bits only for those shift amounts
856 // compatible with its known constraints.
857 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
859 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
861 // If we know the shifter is nonzero, we may be able to infer more known
862 // bits. This check is sunk down as far as possible to avoid the expensive
863 // call to isKnownNonZero if the cheaper checks above fail.
865 if (!ShifterOperandIsNonZero.hasValue())
866 ShifterOperandIsNonZero =
867 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
868 if (*ShifterOperandIsNonZero)
872 Known.Zero &= KZF(Known2.Zero, ShiftAmt);
873 Known.One &= KOF(Known2.One, ShiftAmt);
876 // If there are no compatible shift amounts, then we've proven that the shift
877 // amount must be >= the BitWidth, and the result is undefined. We could
878 // return anything we'd like, but we need to make sure the sets of known bits
879 // stay disjoint (it should be better for some other code to actually
880 // propagate the undef than to pick a value here using known bits).
881 if (Known.Zero.intersects(Known.One))
885 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
886 unsigned Depth, const Query &Q) {
887 unsigned BitWidth = Known.getBitWidth();
889 KnownBits Known2(Known);
890 switch (I->getOpcode()) {
892 case Instruction::Load:
893 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
894 computeKnownBitsFromRangeMetadata(*MD, Known);
896 case Instruction::And: {
897 // If either the LHS or the RHS are Zero, the result is zero.
898 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
899 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
901 // Output known-1 bits are only known if set in both the LHS & RHS.
902 Known.One &= Known2.One;
903 // Output known-0 are known to be clear if zero in either the LHS | RHS.
904 Known.Zero |= Known2.Zero;
906 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
907 // here we handle the more general case of adding any odd number by
908 // matching the form add(x, add(x, y)) where y is odd.
909 // TODO: This could be generalized to clearing any bit set in y where the
910 // following bit is known to be unset in y.
912 if (!Known.Zero[0] && !Known.One[0] &&
913 (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
915 match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
918 computeKnownBits(Y, Known2, Depth + 1, Q);
919 if (Known2.One.countTrailingOnes() > 0)
920 Known.Zero.setBit(0);
924 case Instruction::Or: {
925 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
926 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
928 // Output known-0 bits are only known if clear in both the LHS & RHS.
929 Known.Zero &= Known2.Zero;
930 // Output known-1 are known to be set if set in either the LHS | RHS.
931 Known.One |= Known2.One;
934 case Instruction::Xor: {
935 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
936 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
938 // Output known-0 bits are known if clear or set in both the LHS & RHS.
939 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
940 // Output known-1 are known to be set if set in only one of the LHS, RHS.
941 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
942 Known.Zero = std::move(KnownZeroOut);
945 case Instruction::Mul: {
946 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
947 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
951 case Instruction::UDiv: {
952 // For the purposes of computing leading zeros we can conservatively
953 // treat a udiv as a logical right shift by the power of 2 known to
954 // be less than the denominator.
955 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
956 unsigned LeadZ = Known2.Zero.countLeadingOnes();
959 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
960 unsigned RHSUnknownLeadingOnes = Known2.One.countLeadingZeros();
961 if (RHSUnknownLeadingOnes != BitWidth)
962 LeadZ = std::min(BitWidth,
963 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
965 Known.Zero.setHighBits(LeadZ);
968 case Instruction::Select: {
969 const Value *LHS, *RHS;
970 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
971 if (SelectPatternResult::isMinOrMax(SPF)) {
972 computeKnownBits(RHS, Known, Depth + 1, Q);
973 computeKnownBits(LHS, Known2, Depth + 1, Q);
975 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
976 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
979 unsigned MaxHighOnes = 0;
980 unsigned MaxHighZeros = 0;
981 if (SPF == SPF_SMAX) {
982 // If both sides are negative, the result is negative.
983 if (Known.isNegative() && Known2.isNegative())
984 // We can derive a lower bound on the result by taking the max of the
986 MaxHighOnes = std::max(Known.One.countLeadingOnes(),
987 Known2.One.countLeadingOnes());
988 // If either side is non-negative, the result is non-negative.
989 else if (Known.isNonNegative() || Known2.isNonNegative())
991 } else if (SPF == SPF_SMIN) {
992 // If both sides are non-negative, the result is non-negative.
993 if (Known.isNonNegative() && Known2.isNonNegative())
994 // We can derive an upper bound on the result by taking the max of the
995 // leading zero bits.
996 MaxHighZeros = std::max(Known.Zero.countLeadingOnes(),
997 Known2.Zero.countLeadingOnes());
998 // If either side is negative, the result is negative.
999 else if (Known.isNegative() || Known2.isNegative())
1001 } else if (SPF == SPF_UMAX) {
1002 // We can derive a lower bound on the result by taking the max of the
1003 // leading one bits.
1005 std::max(Known.One.countLeadingOnes(), Known2.One.countLeadingOnes());
1006 } else if (SPF == SPF_UMIN) {
1007 // We can derive an upper bound on the result by taking the max of the
1008 // leading zero bits.
1010 std::max(Known.Zero.countLeadingOnes(), Known2.Zero.countLeadingOnes());
1013 // Only known if known in both the LHS and RHS.
1014 Known.One &= Known2.One;
1015 Known.Zero &= Known2.Zero;
1016 if (MaxHighOnes > 0)
1017 Known.One.setHighBits(MaxHighOnes);
1018 if (MaxHighZeros > 0)
1019 Known.Zero.setHighBits(MaxHighZeros);
1022 case Instruction::FPTrunc:
1023 case Instruction::FPExt:
1024 case Instruction::FPToUI:
1025 case Instruction::FPToSI:
1026 case Instruction::SIToFP:
1027 case Instruction::UIToFP:
1028 break; // Can't work with floating point.
1029 case Instruction::PtrToInt:
1030 case Instruction::IntToPtr:
1031 // Fall through and handle them the same as zext/trunc.
1033 case Instruction::ZExt:
1034 case Instruction::Trunc: {
1035 Type *SrcTy = I->getOperand(0)->getType();
1037 unsigned SrcBitWidth;
1038 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1039 // which fall through here.
1040 SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType());
1042 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1043 Known = Known.zextOrTrunc(SrcBitWidth);
1044 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1045 Known = Known.zextOrTrunc(BitWidth);
1046 // Any top bits are known to be zero.
1047 if (BitWidth > SrcBitWidth)
1048 Known.Zero.setBitsFrom(SrcBitWidth);
1051 case Instruction::BitCast: {
1052 Type *SrcTy = I->getOperand(0)->getType();
1053 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1054 // TODO: For now, not handling conversions like:
1055 // (bitcast i64 %x to <2 x i32>)
1056 !I->getType()->isVectorTy()) {
1057 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1062 case Instruction::SExt: {
1063 // Compute the bits in the result that are not present in the input.
1064 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1066 Known = Known.trunc(SrcBitWidth);
1067 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1068 // If the sign bit of the input is known set or clear, then we know the
1069 // top bits of the result.
1070 Known = Known.sext(BitWidth);
1073 case Instruction::Shl: {
1074 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1075 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1076 auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1077 APInt KZResult = KnownZero << ShiftAmt;
1078 KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1079 // If this shift has "nsw" keyword, then the result is either a poison
1080 // value or has the same sign bit as the first operand.
1081 if (NSW && KnownZero.isSignBitSet())
1082 KZResult.setSignBit();
1086 auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1087 APInt KOResult = KnownOne << ShiftAmt;
1088 if (NSW && KnownOne.isSignBitSet())
1089 KOResult.setSignBit();
1093 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1096 case Instruction::LShr: {
1097 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1098 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1099 APInt KZResult = KnownZero.lshr(ShiftAmt);
1100 // High bits known zero.
1101 KZResult.setHighBits(ShiftAmt);
1105 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1106 return KnownOne.lshr(ShiftAmt);
1109 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1112 case Instruction::AShr: {
1113 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1114 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1115 return KnownZero.ashr(ShiftAmt);
1118 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1119 return KnownOne.ashr(ShiftAmt);
1122 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1125 case Instruction::Sub: {
1126 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1127 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1128 Known, Known2, Depth, Q);
1131 case Instruction::Add: {
1132 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1133 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1134 Known, Known2, Depth, Q);
1137 case Instruction::SRem:
1138 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1139 APInt RA = Rem->getValue().abs();
1140 if (RA.isPowerOf2()) {
1141 APInt LowBits = RA - 1;
1142 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1144 // The low bits of the first operand are unchanged by the srem.
1145 Known.Zero = Known2.Zero & LowBits;
1146 Known.One = Known2.One & LowBits;
1148 // If the first operand is non-negative or has all low bits zero, then
1149 // the upper bits are all zero.
1150 if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1151 Known.Zero |= ~LowBits;
1153 // If the first operand is negative and not all low bits are zero, then
1154 // the upper bits are all one.
1155 if (Known2.isNegative() && LowBits.intersects(Known2.One))
1156 Known.One |= ~LowBits;
1158 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1163 // The sign bit is the LHS's sign bit, except when the result of the
1164 // remainder is zero.
1165 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1166 // If it's known zero, our sign bit is also zero.
1167 if (Known2.isNonNegative())
1168 Known.makeNonNegative();
1171 case Instruction::URem: {
1172 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1173 const APInt &RA = Rem->getValue();
1174 if (RA.isPowerOf2()) {
1175 APInt LowBits = (RA - 1);
1176 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1177 Known.Zero |= ~LowBits;
1178 Known.One &= LowBits;
1183 // Since the result is less than or equal to either operand, any leading
1184 // zero bits in either operand must also exist in the result.
1185 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1186 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1188 unsigned Leaders = std::max(Known.Zero.countLeadingOnes(),
1189 Known2.Zero.countLeadingOnes());
1191 Known.Zero.setHighBits(Leaders);
1195 case Instruction::Alloca: {
1196 const AllocaInst *AI = cast<AllocaInst>(I);
1197 unsigned Align = AI->getAlignment();
1199 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1202 Known.Zero.setLowBits(countTrailingZeros(Align));
1205 case Instruction::GetElementPtr: {
1206 // Analyze all of the subscripts of this getelementptr instruction
1207 // to determine if we can prove known low zero bits.
1208 KnownBits LocalKnown(BitWidth);
1209 computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1210 unsigned TrailZ = LocalKnown.Zero.countTrailingOnes();
1212 gep_type_iterator GTI = gep_type_begin(I);
1213 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1214 Value *Index = I->getOperand(i);
1215 if (StructType *STy = GTI.getStructTypeOrNull()) {
1216 // Handle struct member offset arithmetic.
1218 // Handle case when index is vector zeroinitializer
1219 Constant *CIndex = cast<Constant>(Index);
1220 if (CIndex->isZeroValue())
1223 if (CIndex->getType()->isVectorTy())
1224 Index = CIndex->getSplatValue();
1226 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1227 const StructLayout *SL = Q.DL.getStructLayout(STy);
1228 uint64_t Offset = SL->getElementOffset(Idx);
1229 TrailZ = std::min<unsigned>(TrailZ,
1230 countTrailingZeros(Offset));
1232 // Handle array index arithmetic.
1233 Type *IndexedTy = GTI.getIndexedType();
1234 if (!IndexedTy->isSized()) {
1238 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1239 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1240 LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1241 computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1242 TrailZ = std::min(TrailZ,
1243 unsigned(countTrailingZeros(TypeSize) +
1244 LocalKnown.Zero.countTrailingOnes()));
1248 Known.Zero.setLowBits(TrailZ);
1251 case Instruction::PHI: {
1252 const PHINode *P = cast<PHINode>(I);
1253 // Handle the case of a simple two-predecessor recurrence PHI.
1254 // There's a lot more that could theoretically be done here, but
1255 // this is sufficient to catch some interesting cases.
1256 if (P->getNumIncomingValues() == 2) {
1257 for (unsigned i = 0; i != 2; ++i) {
1258 Value *L = P->getIncomingValue(i);
1259 Value *R = P->getIncomingValue(!i);
1260 Operator *LU = dyn_cast<Operator>(L);
1263 unsigned Opcode = LU->getOpcode();
1264 // Check for operations that have the property that if
1265 // both their operands have low zero bits, the result
1266 // will have low zero bits.
1267 if (Opcode == Instruction::Add ||
1268 Opcode == Instruction::Sub ||
1269 Opcode == Instruction::And ||
1270 Opcode == Instruction::Or ||
1271 Opcode == Instruction::Mul) {
1272 Value *LL = LU->getOperand(0);
1273 Value *LR = LU->getOperand(1);
1274 // Find a recurrence.
1281 // Ok, we have a PHI of the form L op= R. Check for low
1283 computeKnownBits(R, Known2, Depth + 1, Q);
1285 // We need to take the minimum number of known bits
1286 KnownBits Known3(Known);
1287 computeKnownBits(L, Known3, Depth + 1, Q);
1289 Known.Zero.setLowBits(std::min(Known2.Zero.countTrailingOnes(),
1290 Known3.Zero.countTrailingOnes()));
1292 if (DontImproveNonNegativePhiBits)
1295 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1296 if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
1297 // If initial value of recurrence is nonnegative, and we are adding
1298 // a nonnegative number with nsw, the result can only be nonnegative
1299 // or poison value regardless of the number of times we execute the
1300 // add in phi recurrence. If initial value is negative and we are
1301 // adding a negative number with nsw, the result can only be
1302 // negative or poison value. Similar arguments apply to sub and mul.
1304 // (add non-negative, non-negative) --> non-negative
1305 // (add negative, negative) --> negative
1306 if (Opcode == Instruction::Add) {
1307 if (Known2.isNonNegative() && Known3.isNonNegative())
1308 Known.makeNonNegative();
1309 else if (Known2.isNegative() && Known3.isNegative())
1310 Known.makeNegative();
1313 // (sub nsw non-negative, negative) --> non-negative
1314 // (sub nsw negative, non-negative) --> negative
1315 else if (Opcode == Instruction::Sub && LL == I) {
1316 if (Known2.isNonNegative() && Known3.isNegative())
1317 Known.makeNonNegative();
1318 else if (Known2.isNegative() && Known3.isNonNegative())
1319 Known.makeNegative();
1322 // (mul nsw non-negative, non-negative) --> non-negative
1323 else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1324 Known3.isNonNegative())
1325 Known.makeNonNegative();
1333 // Unreachable blocks may have zero-operand PHI nodes.
1334 if (P->getNumIncomingValues() == 0)
1337 // Otherwise take the unions of the known bit sets of the operands,
1338 // taking conservative care to avoid excessive recursion.
1339 if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1340 // Skip if every incoming value references to ourself.
1341 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1344 Known.Zero.setAllBits();
1345 Known.One.setAllBits();
1346 for (Value *IncValue : P->incoming_values()) {
1347 // Skip direct self references.
1348 if (IncValue == P) continue;
1350 Known2 = KnownBits(BitWidth);
1351 // Recurse, but cap the recursion to one level, because we don't
1352 // want to waste time spinning around in loops.
1353 computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1354 Known.Zero &= Known2.Zero;
1355 Known.One &= Known2.One;
1356 // If all bits have been ruled out, there's no need to check
1358 if (!Known.Zero && !Known.One)
1364 case Instruction::Call:
1365 case Instruction::Invoke:
1366 // If range metadata is attached to this call, set known bits from that,
1367 // and then intersect with known bits based on other properties of the
1369 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1370 computeKnownBitsFromRangeMetadata(*MD, Known);
1371 if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1372 computeKnownBits(RV, Known2, Depth + 1, Q);
1373 Known.Zero |= Known2.Zero;
1374 Known.One |= Known2.One;
1376 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1377 switch (II->getIntrinsicID()) {
1379 case Intrinsic::bitreverse:
1380 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1381 Known.Zero |= Known2.Zero.reverseBits();
1382 Known.One |= Known2.One.reverseBits();
1384 case Intrinsic::bswap:
1385 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1386 Known.Zero |= Known2.Zero.byteSwap();
1387 Known.One |= Known2.One.byteSwap();
1389 case Intrinsic::ctlz:
1390 case Intrinsic::cttz: {
1391 unsigned LowBits = Log2_32(BitWidth)+1;
1392 // If this call is undefined for 0, the result will be less than 2^n.
1393 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1395 Known.Zero.setBitsFrom(LowBits);
1398 case Intrinsic::ctpop: {
1399 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1400 // We can bound the space the count needs. Also, bits known to be zero
1401 // can't contribute to the population.
1402 unsigned BitsPossiblySet = BitWidth - Known2.Zero.countPopulation();
1403 unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1404 Known.Zero.setBitsFrom(LowBits);
1405 // TODO: we could bound KnownOne using the lower bound on the number
1406 // of bits which might be set provided by popcnt KnownOne2.
1409 case Intrinsic::x86_sse42_crc32_64_64:
1410 Known.Zero.setBitsFrom(32);
1415 case Instruction::ExtractElement:
1416 // Look through extract element. At the moment we keep this simple and skip
1417 // tracking the specific element. But at least we might find information
1418 // valid for all elements of the vector (for example if vector is sign
1419 // extended, shifted, etc).
1420 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1422 case Instruction::ExtractValue:
1423 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1424 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1425 if (EVI->getNumIndices() != 1) break;
1426 if (EVI->getIndices()[0] == 0) {
1427 switch (II->getIntrinsicID()) {
1429 case Intrinsic::uadd_with_overflow:
1430 case Intrinsic::sadd_with_overflow:
1431 computeKnownBitsAddSub(true, II->getArgOperand(0),
1432 II->getArgOperand(1), false, Known, Known2,
1435 case Intrinsic::usub_with_overflow:
1436 case Intrinsic::ssub_with_overflow:
1437 computeKnownBitsAddSub(false, II->getArgOperand(0),
1438 II->getArgOperand(1), false, Known, Known2,
1441 case Intrinsic::umul_with_overflow:
1442 case Intrinsic::smul_with_overflow:
1443 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1444 Known, Known2, Depth, Q);
1452 /// Determine which bits of V are known to be either zero or one and return
1453 /// them in the Known bit set.
1455 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1456 /// we cannot optimize based on the assumption that it is zero without changing
1457 /// it to be an explicit zero. If we don't change it to zero, other code could
1458 /// optimized based on the contradictory assumption that it is non-zero.
1459 /// Because instcombine aggressively folds operations with undef args anyway,
1460 /// this won't lose us code quality.
1462 /// This function is defined on values with integer type, values with pointer
1463 /// type, and vectors of integers. In the case
1464 /// where V is a vector, known zero, and known one values are the
1465 /// same width as the vector element, and the bit is set only if it is true
1466 /// for all of the elements in the vector.
1467 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1469 assert(V && "No Value?");
1470 assert(Depth <= MaxDepth && "Limit Search Depth");
1471 unsigned BitWidth = Known.getBitWidth();
1473 assert((V->getType()->isIntOrIntVectorTy() ||
1474 V->getType()->getScalarType()->isPointerTy()) &&
1475 "Not integer or pointer type!");
1476 assert((Q.DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1477 (!V->getType()->isIntOrIntVectorTy() ||
1478 V->getType()->getScalarSizeInBits() == BitWidth) &&
1479 "V and Known should have same BitWidth");
1483 if (match(V, m_APInt(C))) {
1484 // We know all of the bits for a scalar constant or a splat vector constant!
1486 Known.Zero = ~Known.One;
1489 // Null and aggregate-zero are all-zeros.
1490 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1494 // Handle a constant vector by taking the intersection of the known bits of
1496 if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1497 // We know that CDS must be a vector of integers. Take the intersection of
1499 Known.Zero.setAllBits(); Known.One.setAllBits();
1500 APInt Elt(BitWidth, 0);
1501 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1502 Elt = CDS->getElementAsInteger(i);
1509 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1510 // We know that CV must be a vector of integers. Take the intersection of
1512 Known.Zero.setAllBits(); Known.One.setAllBits();
1513 APInt Elt(BitWidth, 0);
1514 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1515 Constant *Element = CV->getAggregateElement(i);
1516 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1521 Elt = ElementCI->getValue();
1528 // Start out not knowing anything.
1531 // We can't imply anything about undefs.
1532 if (isa<UndefValue>(V))
1535 // There's no point in looking through other users of ConstantData for
1536 // assumptions. Confirm that we've handled them all.
1537 assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1539 // Limit search depth.
1540 // All recursive calls that increase depth must come after this.
1541 if (Depth == MaxDepth)
1544 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1545 // the bits of its aliasee.
1546 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1547 if (!GA->isInterposable())
1548 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1552 if (const Operator *I = dyn_cast<Operator>(V))
1553 computeKnownBitsFromOperator(I, Known, Depth, Q);
1555 // Aligned pointers have trailing zeros - refine Known.Zero set
1556 if (V->getType()->isPointerTy()) {
1557 unsigned Align = V->getPointerAlignment(Q.DL);
1559 Known.Zero.setLowBits(countTrailingZeros(Align));
1562 // computeKnownBitsFromAssume strictly refines Known.
1563 // Therefore, we run them after computeKnownBitsFromOperator.
1565 // Check whether a nearby assume intrinsic can determine some known bits.
1566 computeKnownBitsFromAssume(V, Known, Depth, Q);
1568 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1571 /// Determine whether the sign bit is known to be zero or one.
1572 /// Convenience wrapper around computeKnownBits.
1573 void ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
1574 unsigned Depth, const Query &Q) {
1575 KnownBits Bits(getBitWidth(V->getType(), Q.DL));
1576 computeKnownBits(V, Bits, Depth, Q);
1577 KnownOne = Bits.isNegative();
1578 KnownZero = Bits.isNonNegative();
1581 /// Return true if the given value is known to have exactly one
1582 /// bit set when defined. For vectors return true if every element is known to
1583 /// be a power of two when defined. Supports values with integer or pointer
1584 /// types and vectors of integers.
1585 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1587 if (const Constant *C = dyn_cast<Constant>(V)) {
1588 if (C->isNullValue())
1591 const APInt *ConstIntOrConstSplatInt;
1592 if (match(C, m_APInt(ConstIntOrConstSplatInt)))
1593 return ConstIntOrConstSplatInt->isPowerOf2();
1596 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1597 // it is shifted off the end then the result is undefined.
1598 if (match(V, m_Shl(m_One(), m_Value())))
1601 // (signmask) >>l X is clearly a power of two if the one is not shifted off
1602 // the bottom. If it is shifted off the bottom then the result is undefined.
1603 if (match(V, m_LShr(m_SignMask(), m_Value())))
1606 // The remaining tests are all recursive, so bail out if we hit the limit.
1607 if (Depth++ == MaxDepth)
1610 Value *X = nullptr, *Y = nullptr;
1611 // A shift left or a logical shift right of a power of two is a power of two
1613 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1614 match(V, m_LShr(m_Value(X), m_Value()))))
1615 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1617 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1618 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1620 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1621 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1622 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1624 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1625 // A power of two and'd with anything is a power of two or zero.
1626 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1627 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1629 // X & (-X) is always a power of two or zero.
1630 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1635 // Adding a power-of-two or zero to the same power-of-two or zero yields
1636 // either the original power-of-two, a larger power-of-two or zero.
1637 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1638 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1639 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1640 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1641 match(X, m_And(m_Value(), m_Specific(Y))))
1642 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1644 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1645 match(Y, m_And(m_Value(), m_Specific(X))))
1646 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1649 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1650 KnownBits LHSBits(BitWidth);
1651 computeKnownBits(X, LHSBits, Depth, Q);
1653 KnownBits RHSBits(BitWidth);
1654 computeKnownBits(Y, RHSBits, Depth, Q);
1655 // If i8 V is a power of two or zero:
1656 // ZeroBits: 1 1 1 0 1 1 1 1
1657 // ~ZeroBits: 0 0 0 1 0 0 0 0
1658 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1659 // If OrZero isn't set, we cannot give back a zero result.
1660 // Make sure either the LHS or RHS has a bit set.
1661 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1666 // An exact divide or right shift can only shift off zero bits, so the result
1667 // is a power of two only if the first operand is a power of two and not
1668 // copying a sign bit (sdiv int_min, 2).
1669 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1670 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1671 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1678 /// \brief Test whether a GEP's result is known to be non-null.
1680 /// Uses properties inherent in a GEP to try to determine whether it is known
1683 /// Currently this routine does not support vector GEPs.
1684 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1686 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1689 // FIXME: Support vector-GEPs.
1690 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1692 // If the base pointer is non-null, we cannot walk to a null address with an
1693 // inbounds GEP in address space zero.
1694 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1697 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1698 // If so, then the GEP cannot produce a null pointer, as doing so would
1699 // inherently violate the inbounds contract within address space zero.
1700 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1701 GTI != GTE; ++GTI) {
1702 // Struct types are easy -- they must always be indexed by a constant.
1703 if (StructType *STy = GTI.getStructTypeOrNull()) {
1704 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1705 unsigned ElementIdx = OpC->getZExtValue();
1706 const StructLayout *SL = Q.DL.getStructLayout(STy);
1707 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1708 if (ElementOffset > 0)
1713 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1714 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1717 // Fast path the constant operand case both for efficiency and so we don't
1718 // increment Depth when just zipping down an all-constant GEP.
1719 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1725 // We post-increment Depth here because while isKnownNonZero increments it
1726 // as well, when we pop back up that increment won't persist. We don't want
1727 // to recurse 10k times just because we have 10k GEP operands. We don't
1728 // bail completely out because we want to handle constant GEPs regardless
1730 if (Depth++ >= MaxDepth)
1733 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1740 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1741 /// ensure that the value it's attached to is never Value? 'RangeType' is
1742 /// is the type of the value described by the range.
1743 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1744 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1745 assert(NumRanges >= 1);
1746 for (unsigned i = 0; i < NumRanges; ++i) {
1747 ConstantInt *Lower =
1748 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1749 ConstantInt *Upper =
1750 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1751 ConstantRange Range(Lower->getValue(), Upper->getValue());
1752 if (Range.contains(Value))
1758 /// Return true if the given value is known to be non-zero when defined. For
1759 /// vectors, return true if every element is known to be non-zero when
1760 /// defined. For pointers, if the context instruction and dominator tree are
1761 /// specified, perform context-sensitive analysis and return true if the
1762 /// pointer couldn't possibly be null at the specified instruction.
1763 /// Supports values with integer or pointer type and vectors of integers.
1764 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1765 if (auto *C = dyn_cast<Constant>(V)) {
1766 if (C->isNullValue())
1768 if (isa<ConstantInt>(C))
1769 // Must be non-zero due to null test above.
1772 // For constant vectors, check that all elements are undefined or known
1773 // non-zero to determine that the whole vector is known non-zero.
1774 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1775 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1776 Constant *Elt = C->getAggregateElement(i);
1777 if (!Elt || Elt->isNullValue())
1779 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1788 if (auto *I = dyn_cast<Instruction>(V)) {
1789 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1790 // If the possible ranges don't contain zero, then the value is
1791 // definitely non-zero.
1792 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1793 const APInt ZeroValue(Ty->getBitWidth(), 0);
1794 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1800 // The remaining tests are all recursive, so bail out if we hit the limit.
1801 if (Depth++ >= MaxDepth)
1804 // Check for pointer simplifications.
1805 if (V->getType()->isPointerTy()) {
1806 if (isKnownNonNullAt(V, Q.CxtI, Q.DT))
1808 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1809 if (isGEPKnownNonNull(GEP, Depth, Q))
1813 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1815 // X | Y != 0 if X != 0 or Y != 0.
1816 Value *X = nullptr, *Y = nullptr;
1817 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1818 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1820 // ext X != 0 if X != 0.
1821 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1822 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
1824 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1825 // if the lowest bit is shifted off the end.
1826 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1827 // shl nuw can't remove any non-zero bits.
1828 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1829 if (BO->hasNoUnsignedWrap())
1830 return isKnownNonZero(X, Depth, Q);
1832 KnownBits Known(BitWidth);
1833 computeKnownBits(X, Known, Depth, Q);
1837 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1838 // defined if the sign bit is shifted off the end.
1839 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1840 // shr exact can only shift out zero bits.
1841 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1843 return isKnownNonZero(X, Depth, Q);
1845 bool XKnownNonNegative, XKnownNegative;
1846 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1850 // If the shifter operand is a constant, and all of the bits shifted
1851 // out are known to be zero, and X is known non-zero then at least one
1852 // non-zero bit must remain.
1853 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1854 KnownBits Known(BitWidth);
1855 computeKnownBits(X, Known, Depth, Q);
1857 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1858 // Is there a known one in the portion not shifted out?
1859 if (Known.One.countLeadingZeros() < BitWidth - ShiftVal)
1861 // Are all the bits to be shifted out known zero?
1862 if (Known.Zero.countTrailingOnes() >= ShiftVal)
1863 return isKnownNonZero(X, Depth, Q);
1866 // div exact can only produce a zero if the dividend is zero.
1867 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1868 return isKnownNonZero(X, Depth, Q);
1871 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1872 bool XKnownNonNegative, XKnownNegative;
1873 bool YKnownNonNegative, YKnownNegative;
1874 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1875 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Depth, Q);
1877 // If X and Y are both non-negative (as signed values) then their sum is not
1878 // zero unless both X and Y are zero.
1879 if (XKnownNonNegative && YKnownNonNegative)
1880 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
1883 // If X and Y are both negative (as signed values) then their sum is not
1884 // zero unless both X and Y equal INT_MIN.
1885 if (XKnownNegative && YKnownNegative) {
1886 KnownBits Known(BitWidth);
1887 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1888 // The sign bit of X is set. If some other bit is set then X is not equal
1890 computeKnownBits(X, Known, Depth, Q);
1891 if (Known.One.intersects(Mask))
1893 // The sign bit of Y is set. If some other bit is set then Y is not equal
1895 computeKnownBits(Y, Known, Depth, Q);
1896 if (Known.One.intersects(Mask))
1900 // The sum of a non-negative number and a power of two is not zero.
1901 if (XKnownNonNegative &&
1902 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
1904 if (YKnownNonNegative &&
1905 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
1909 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1910 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1911 // If X and Y are non-zero then so is X * Y as long as the multiplication
1912 // does not overflow.
1913 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1914 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
1917 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1918 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
1919 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
1920 isKnownNonZero(SI->getFalseValue(), Depth, Q))
1924 else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
1925 // Try and detect a recurrence that monotonically increases from a
1926 // starting value, as these are common as induction variables.
1927 if (PN->getNumIncomingValues() == 2) {
1928 Value *Start = PN->getIncomingValue(0);
1929 Value *Induction = PN->getIncomingValue(1);
1930 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1931 std::swap(Start, Induction);
1932 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1933 if (!C->isZero() && !C->isNegative()) {
1935 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1936 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1942 // Check if all incoming values are non-zero constant.
1943 bool AllNonZeroConstants = all_of(PN->operands(), [](Value *V) {
1944 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZeroValue();
1946 if (AllNonZeroConstants)
1950 KnownBits Known(BitWidth);
1951 computeKnownBits(V, Known, Depth, Q);
1952 return Known.One != 0;
1955 /// Return true if V2 == V1 + X, where X is known non-zero.
1956 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
1957 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
1958 if (!BO || BO->getOpcode() != Instruction::Add)
1960 Value *Op = nullptr;
1961 if (V2 == BO->getOperand(0))
1962 Op = BO->getOperand(1);
1963 else if (V2 == BO->getOperand(1))
1964 Op = BO->getOperand(0);
1967 return isKnownNonZero(Op, 0, Q);
1970 /// Return true if it is known that V1 != V2.
1971 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
1972 if (V1->getType()->isVectorTy() || V1 == V2)
1974 if (V1->getType() != V2->getType())
1975 // We can't look through casts yet.
1977 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
1980 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
1981 // Are any known bits in V1 contradictory to known bits in V2? If V1
1982 // has a known zero where V2 has a known one, they must not be equal.
1983 auto BitWidth = Ty->getBitWidth();
1984 KnownBits Known1(BitWidth);
1985 computeKnownBits(V1, Known1, 0, Q);
1986 KnownBits Known2(BitWidth);
1987 computeKnownBits(V2, Known2, 0, Q);
1989 APInt OppositeBits = (Known1.Zero & Known2.One) |
1990 (Known2.Zero & Known1.One);
1991 if (OppositeBits.getBoolValue())
1997 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1998 /// simplify operations downstream. Mask is known to be zero for bits that V
2001 /// This function is defined on values with integer type, values with pointer
2002 /// type, and vectors of integers. In the case
2003 /// where V is a vector, the mask, known zero, and known one values are the
2004 /// same width as the vector element, and the bit is set only if it is true
2005 /// for all of the elements in the vector.
2006 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2008 KnownBits Known(Mask.getBitWidth());
2009 computeKnownBits(V, Known, Depth, Q);
2010 return Mask.isSubsetOf(Known.Zero);
2013 /// For vector constants, loop over the elements and find the constant with the
2014 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2015 /// or if any element was not analyzed; otherwise, return the count for the
2016 /// element with the minimum number of sign bits.
2017 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2019 const auto *CV = dyn_cast<Constant>(V);
2020 if (!CV || !CV->getType()->isVectorTy())
2023 unsigned MinSignBits = TyBits;
2024 unsigned NumElts = CV->getType()->getVectorNumElements();
2025 for (unsigned i = 0; i != NumElts; ++i) {
2026 // If we find a non-ConstantInt, bail out.
2027 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2031 // If the sign bit is 1, flip the bits, so we always count leading zeros.
2032 APInt EltVal = Elt->getValue();
2033 if (EltVal.isNegative())
2035 MinSignBits = std::min(MinSignBits, EltVal.countLeadingZeros());
2041 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2044 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2046 unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2047 assert(Result > 0 && "At least one sign bit needs to be present!");
2051 /// Return the number of times the sign bit of the register is replicated into
2052 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2053 /// (itself), but other cases can give us information. For example, immediately
2054 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2055 /// other, so we return 3. For vectors, return the number of sign bits for the
2056 /// vector element with the mininum number of known sign bits.
2057 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2060 // We return the minimum number of sign bits that are guaranteed to be present
2061 // in V, so for undef we have to conservatively return 1. We don't have the
2062 // same behavior for poison though -- that's a FIXME today.
2064 unsigned TyBits = Q.DL.getTypeSizeInBits(V->getType()->getScalarType());
2066 unsigned FirstAnswer = 1;
2068 // Note that ConstantInt is handled by the general computeKnownBits case
2071 if (Depth == MaxDepth)
2072 return 1; // Limit search depth.
2074 const Operator *U = dyn_cast<Operator>(V);
2075 switch (Operator::getOpcode(V)) {
2077 case Instruction::SExt:
2078 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2079 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2081 case Instruction::SDiv: {
2082 const APInt *Denominator;
2083 // sdiv X, C -> adds log(C) sign bits.
2084 if (match(U->getOperand(1), m_APInt(Denominator))) {
2086 // Ignore non-positive denominator.
2087 if (!Denominator->isStrictlyPositive())
2090 // Calculate the incoming numerator bits.
2091 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2093 // Add floor(log(C)) bits to the numerator bits.
2094 return std::min(TyBits, NumBits + Denominator->logBase2());
2099 case Instruction::SRem: {
2100 const APInt *Denominator;
2101 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2102 // positive constant. This let us put a lower bound on the number of sign
2104 if (match(U->getOperand(1), m_APInt(Denominator))) {
2106 // Ignore non-positive denominator.
2107 if (!Denominator->isStrictlyPositive())
2110 // Calculate the incoming numerator bits. SRem by a positive constant
2111 // can't lower the number of sign bits.
2113 ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2115 // Calculate the leading sign bit constraints by examining the
2116 // denominator. Given that the denominator is positive, there are two
2119 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2120 // (1 << ceilLogBase2(C)).
2122 // 2. the numerator is negative. Then the result range is (-C,0] and
2123 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2125 // Thus a lower bound on the number of sign bits is `TyBits -
2126 // ceilLogBase2(C)`.
2128 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2129 return std::max(NumrBits, ResBits);
2134 case Instruction::AShr: {
2135 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2136 // ashr X, C -> adds C sign bits. Vectors too.
2138 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2139 unsigned ShAmtLimited = ShAmt->getZExtValue();
2140 if (ShAmtLimited >= TyBits)
2141 break; // Bad shift.
2142 Tmp += ShAmtLimited;
2143 if (Tmp > TyBits) Tmp = TyBits;
2147 case Instruction::Shl: {
2149 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2150 // shl destroys sign bits.
2151 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2152 Tmp2 = ShAmt->getZExtValue();
2153 if (Tmp2 >= TyBits || // Bad shift.
2154 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2159 case Instruction::And:
2160 case Instruction::Or:
2161 case Instruction::Xor: // NOT is handled here.
2162 // Logical binary ops preserve the number of sign bits at the worst.
2163 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2165 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2166 FirstAnswer = std::min(Tmp, Tmp2);
2167 // We computed what we know about the sign bits as our first
2168 // answer. Now proceed to the generic code that uses
2169 // computeKnownBits, and pick whichever answer is better.
2173 case Instruction::Select:
2174 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2175 if (Tmp == 1) return 1; // Early out.
2176 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2177 return std::min(Tmp, Tmp2);
2179 case Instruction::Add:
2180 // Add can have at most one carry bit. Thus we know that the output
2181 // is, at worst, one more bit than the inputs.
2182 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2183 if (Tmp == 1) return 1; // Early out.
2185 // Special case decrementing a value (ADD X, -1):
2186 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2187 if (CRHS->isAllOnesValue()) {
2188 KnownBits Known(TyBits);
2189 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2191 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2193 if ((Known.Zero | 1).isAllOnesValue())
2196 // If we are subtracting one from a positive number, there is no carry
2197 // out of the result.
2198 if (Known.isNonNegative())
2202 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2203 if (Tmp2 == 1) return 1;
2204 return std::min(Tmp, Tmp2)-1;
2206 case Instruction::Sub:
2207 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2208 if (Tmp2 == 1) return 1;
2211 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2212 if (CLHS->isNullValue()) {
2213 KnownBits Known(TyBits);
2214 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2215 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2217 if ((Known.Zero | 1).isAllOnesValue())
2220 // If the input is known to be positive (the sign bit is known clear),
2221 // the output of the NEG has the same number of sign bits as the input.
2222 if (Known.isNonNegative())
2225 // Otherwise, we treat this like a SUB.
2228 // Sub can have at most one carry bit. Thus we know that the output
2229 // is, at worst, one more bit than the inputs.
2230 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2231 if (Tmp == 1) return 1; // Early out.
2232 return std::min(Tmp, Tmp2)-1;
2234 case Instruction::PHI: {
2235 const PHINode *PN = cast<PHINode>(U);
2236 unsigned NumIncomingValues = PN->getNumIncomingValues();
2237 // Don't analyze large in-degree PHIs.
2238 if (NumIncomingValues > 4) break;
2239 // Unreachable blocks may have zero-operand PHI nodes.
2240 if (NumIncomingValues == 0) break;
2242 // Take the minimum of all incoming values. This can't infinitely loop
2243 // because of our depth threshold.
2244 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2245 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2246 if (Tmp == 1) return Tmp;
2248 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2253 case Instruction::Trunc:
2254 // FIXME: it's tricky to do anything useful for this, but it is an important
2255 // case for targets like X86.
2258 case Instruction::ExtractElement:
2259 // Look through extract element. At the moment we keep this simple and skip
2260 // tracking the specific element. But at least we might find information
2261 // valid for all elements of the vector (for example if vector is sign
2262 // extended, shifted, etc).
2263 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2266 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2267 // use this information.
2269 // If we can examine all elements of a vector constant successfully, we're
2270 // done (we can't do any better than that). If not, keep trying.
2271 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2274 KnownBits Known(TyBits);
2275 computeKnownBits(V, Known, Depth, Q);
2277 // If we know that the sign bit is either zero or one, determine the number of
2278 // identical bits in the top of the input value.
2279 if (Known.isNonNegative())
2280 return std::max(FirstAnswer, Known.Zero.countLeadingOnes());
2282 if (Known.isNegative())
2283 return std::max(FirstAnswer, Known.One.countLeadingOnes());
2285 // computeKnownBits gave us no extra information about the top bits.
2289 /// This function computes the integer multiple of Base that equals V.
2290 /// If successful, it returns true and returns the multiple in
2291 /// Multiple. If unsuccessful, it returns false. It looks
2292 /// through SExt instructions only if LookThroughSExt is true.
2293 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2294 bool LookThroughSExt, unsigned Depth) {
2295 const unsigned MaxDepth = 6;
2297 assert(V && "No Value?");
2298 assert(Depth <= MaxDepth && "Limit Search Depth");
2299 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2301 Type *T = V->getType();
2303 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2313 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2314 Constant *BaseVal = ConstantInt::get(T, Base);
2315 if (CO && CO == BaseVal) {
2317 Multiple = ConstantInt::get(T, 1);
2321 if (CI && CI->getZExtValue() % Base == 0) {
2322 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2326 if (Depth == MaxDepth) return false; // Limit search depth.
2328 Operator *I = dyn_cast<Operator>(V);
2329 if (!I) return false;
2331 switch (I->getOpcode()) {
2333 case Instruction::SExt:
2334 if (!LookThroughSExt) return false;
2335 // otherwise fall through to ZExt
2336 case Instruction::ZExt:
2337 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2338 LookThroughSExt, Depth+1);
2339 case Instruction::Shl:
2340 case Instruction::Mul: {
2341 Value *Op0 = I->getOperand(0);
2342 Value *Op1 = I->getOperand(1);
2344 if (I->getOpcode() == Instruction::Shl) {
2345 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2346 if (!Op1CI) return false;
2347 // Turn Op0 << Op1 into Op0 * 2^Op1
2348 APInt Op1Int = Op1CI->getValue();
2349 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2350 APInt API(Op1Int.getBitWidth(), 0);
2351 API.setBit(BitToSet);
2352 Op1 = ConstantInt::get(V->getContext(), API);
2355 Value *Mul0 = nullptr;
2356 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2357 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2358 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2359 if (Op1C->getType()->getPrimitiveSizeInBits() <
2360 MulC->getType()->getPrimitiveSizeInBits())
2361 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2362 if (Op1C->getType()->getPrimitiveSizeInBits() >
2363 MulC->getType()->getPrimitiveSizeInBits())
2364 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2366 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2367 Multiple = ConstantExpr::getMul(MulC, Op1C);
2371 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2372 if (Mul0CI->getValue() == 1) {
2373 // V == Base * Op1, so return Op1
2379 Value *Mul1 = nullptr;
2380 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2381 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2382 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2383 if (Op0C->getType()->getPrimitiveSizeInBits() <
2384 MulC->getType()->getPrimitiveSizeInBits())
2385 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2386 if (Op0C->getType()->getPrimitiveSizeInBits() >
2387 MulC->getType()->getPrimitiveSizeInBits())
2388 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2390 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2391 Multiple = ConstantExpr::getMul(MulC, Op0C);
2395 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2396 if (Mul1CI->getValue() == 1) {
2397 // V == Base * Op0, so return Op0
2405 // We could not determine if V is a multiple of Base.
2409 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2410 const TargetLibraryInfo *TLI) {
2411 const Function *F = ICS.getCalledFunction();
2413 return Intrinsic::not_intrinsic;
2415 if (F->isIntrinsic())
2416 return F->getIntrinsicID();
2419 return Intrinsic::not_intrinsic;
2422 // We're going to make assumptions on the semantics of the functions, check
2423 // that the target knows that it's available in this environment and it does
2424 // not have local linkage.
2425 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2426 return Intrinsic::not_intrinsic;
2428 if (!ICS.onlyReadsMemory())
2429 return Intrinsic::not_intrinsic;
2431 // Otherwise check if we have a call to a function that can be turned into a
2432 // vector intrinsic.
2439 return Intrinsic::sin;
2443 return Intrinsic::cos;
2447 return Intrinsic::exp;
2451 return Intrinsic::exp2;
2455 return Intrinsic::log;
2457 case LibFunc_log10f:
2458 case LibFunc_log10l:
2459 return Intrinsic::log10;
2463 return Intrinsic::log2;
2467 return Intrinsic::fabs;
2471 return Intrinsic::minnum;
2475 return Intrinsic::maxnum;
2476 case LibFunc_copysign:
2477 case LibFunc_copysignf:
2478 case LibFunc_copysignl:
2479 return Intrinsic::copysign;
2481 case LibFunc_floorf:
2482 case LibFunc_floorl:
2483 return Intrinsic::floor;
2487 return Intrinsic::ceil;
2489 case LibFunc_truncf:
2490 case LibFunc_truncl:
2491 return Intrinsic::trunc;
2495 return Intrinsic::rint;
2496 case LibFunc_nearbyint:
2497 case LibFunc_nearbyintf:
2498 case LibFunc_nearbyintl:
2499 return Intrinsic::nearbyint;
2501 case LibFunc_roundf:
2502 case LibFunc_roundl:
2503 return Intrinsic::round;
2507 return Intrinsic::pow;
2511 if (ICS->hasNoNaNs())
2512 return Intrinsic::sqrt;
2513 return Intrinsic::not_intrinsic;
2516 return Intrinsic::not_intrinsic;
2519 /// Return true if we can prove that the specified FP value is never equal to
2522 /// NOTE: this function will need to be revisited when we support non-default
2525 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2527 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2528 return !CFP->getValueAPF().isNegZero();
2530 if (Depth == MaxDepth)
2531 return false; // Limit search depth.
2533 const Operator *I = dyn_cast<Operator>(V);
2534 if (!I) return false;
2536 // Check if the nsz fast-math flag is set
2537 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2538 if (FPO->hasNoSignedZeros())
2541 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2542 if (I->getOpcode() == Instruction::FAdd)
2543 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2544 if (CFP->isNullValue())
2547 // sitofp and uitofp turn into +0.0 for zero.
2548 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2551 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2552 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2556 // sqrt(-0.0) = -0.0, no other negative results are possible.
2557 case Intrinsic::sqrt:
2558 return CannotBeNegativeZero(CI->getArgOperand(0), TLI, Depth + 1);
2560 case Intrinsic::fabs:
2568 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2569 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2570 /// bit despite comparing equal.
2571 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
2572 const TargetLibraryInfo *TLI,
2575 // TODO: This function does not do the right thing when SignBitOnly is true
2576 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2577 // which flips the sign bits of NaNs. See
2578 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2580 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2581 return !CFP->getValueAPF().isNegative() ||
2582 (!SignBitOnly && CFP->getValueAPF().isZero());
2585 if (Depth == MaxDepth)
2586 return false; // Limit search depth.
2588 const Operator *I = dyn_cast<Operator>(V);
2592 switch (I->getOpcode()) {
2595 // Unsigned integers are always nonnegative.
2596 case Instruction::UIToFP:
2598 case Instruction::FMul:
2599 // x*x is always non-negative or a NaN.
2600 if (I->getOperand(0) == I->getOperand(1) &&
2601 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2605 case Instruction::FAdd:
2606 case Instruction::FDiv:
2607 case Instruction::FRem:
2608 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2610 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2612 case Instruction::Select:
2613 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2615 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2617 case Instruction::FPExt:
2618 case Instruction::FPTrunc:
2619 // Widening/narrowing never change sign.
2620 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2622 case Instruction::Call:
2623 const auto *CI = cast<CallInst>(I);
2624 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2628 case Intrinsic::maxnum:
2629 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2631 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2633 case Intrinsic::minnum:
2634 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2636 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2638 case Intrinsic::exp:
2639 case Intrinsic::exp2:
2640 case Intrinsic::fabs:
2643 case Intrinsic::sqrt:
2644 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2647 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2648 CannotBeNegativeZero(CI->getOperand(0), TLI));
2650 case Intrinsic::powi:
2651 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2652 // powi(x,n) is non-negative if n is even.
2653 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2656 // TODO: This is not correct. Given that exp is an integer, here are the
2657 // ways that pow can return a negative value:
2659 // pow(x, exp) --> negative if exp is odd and x is negative.
2660 // pow(-0, exp) --> -inf if exp is negative odd.
2661 // pow(-0, exp) --> -0 if exp is positive odd.
2662 // pow(-inf, exp) --> -0 if exp is negative odd.
2663 // pow(-inf, exp) --> -inf if exp is positive odd.
2665 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2666 // but we must return false if x == -0. Unfortunately we do not currently
2667 // have a way of expressing this constraint. See details in
2668 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2669 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2672 case Intrinsic::fma:
2673 case Intrinsic::fmuladd:
2674 // x*x+y is non-negative if y is non-negative.
2675 return I->getOperand(0) == I->getOperand(1) &&
2676 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2677 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2685 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2686 const TargetLibraryInfo *TLI) {
2687 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
2690 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
2691 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
2694 /// If the specified value can be set by repeating the same byte in memory,
2695 /// return the i8 value that it is represented with. This is
2696 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2697 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2698 /// byte store (e.g. i16 0x1234), return null.
2699 Value *llvm::isBytewiseValue(Value *V) {
2700 // All byte-wide stores are splatable, even of arbitrary variables.
2701 if (V->getType()->isIntegerTy(8)) return V;
2703 // Handle 'null' ConstantArrayZero etc.
2704 if (Constant *C = dyn_cast<Constant>(V))
2705 if (C->isNullValue())
2706 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2708 // Constant float and double values can be handled as integer values if the
2709 // corresponding integer value is "byteable". An important case is 0.0.
2710 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2711 if (CFP->getType()->isFloatTy())
2712 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2713 if (CFP->getType()->isDoubleTy())
2714 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2715 // Don't handle long double formats, which have strange constraints.
2718 // We can handle constant integers that are multiple of 8 bits.
2719 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2720 if (CI->getBitWidth() % 8 == 0) {
2721 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2723 if (!CI->getValue().isSplat(8))
2725 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2729 // A ConstantDataArray/Vector is splatable if all its members are equal and
2731 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2732 Value *Elt = CA->getElementAsConstant(0);
2733 Value *Val = isBytewiseValue(Elt);
2737 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2738 if (CA->getElementAsConstant(I) != Elt)
2744 // Conceptually, we could handle things like:
2745 // %a = zext i8 %X to i16
2746 // %b = shl i16 %a, 8
2747 // %c = or i16 %a, %b
2748 // but until there is an example that actually needs this, it doesn't seem
2749 // worth worrying about.
2754 // This is the recursive version of BuildSubAggregate. It takes a few different
2755 // arguments. Idxs is the index within the nested struct From that we are
2756 // looking at now (which is of type IndexedType). IdxSkip is the number of
2757 // indices from Idxs that should be left out when inserting into the resulting
2758 // struct. To is the result struct built so far, new insertvalue instructions
2760 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2761 SmallVectorImpl<unsigned> &Idxs,
2763 Instruction *InsertBefore) {
2764 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2766 // Save the original To argument so we can modify it
2768 // General case, the type indexed by Idxs is a struct
2769 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2770 // Process each struct element recursively
2773 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2777 // Couldn't find any inserted value for this index? Cleanup
2778 while (PrevTo != OrigTo) {
2779 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2780 PrevTo = Del->getAggregateOperand();
2781 Del->eraseFromParent();
2783 // Stop processing elements
2787 // If we successfully found a value for each of our subaggregates
2791 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2792 // the struct's elements had a value that was inserted directly. In the latter
2793 // case, perhaps we can't determine each of the subelements individually, but
2794 // we might be able to find the complete struct somewhere.
2796 // Find the value that is at that particular spot
2797 Value *V = FindInsertedValue(From, Idxs);
2802 // Insert the value in the new (sub) aggregrate
2803 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2804 "tmp", InsertBefore);
2807 // This helper takes a nested struct and extracts a part of it (which is again a
2808 // struct) into a new value. For example, given the struct:
2809 // { a, { b, { c, d }, e } }
2810 // and the indices "1, 1" this returns
2813 // It does this by inserting an insertvalue for each element in the resulting
2814 // struct, as opposed to just inserting a single struct. This will only work if
2815 // each of the elements of the substruct are known (ie, inserted into From by an
2816 // insertvalue instruction somewhere).
2818 // All inserted insertvalue instructions are inserted before InsertBefore
2819 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2820 Instruction *InsertBefore) {
2821 assert(InsertBefore && "Must have someplace to insert!");
2822 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2824 Value *To = UndefValue::get(IndexedType);
2825 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2826 unsigned IdxSkip = Idxs.size();
2828 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2831 /// Given an aggregrate and an sequence of indices, see if
2832 /// the scalar value indexed is already around as a register, for example if it
2833 /// were inserted directly into the aggregrate.
2835 /// If InsertBefore is not null, this function will duplicate (modified)
2836 /// insertvalues when a part of a nested struct is extracted.
2837 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2838 Instruction *InsertBefore) {
2839 // Nothing to index? Just return V then (this is useful at the end of our
2841 if (idx_range.empty())
2843 // We have indices, so V should have an indexable type.
2844 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2845 "Not looking at a struct or array?");
2846 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2847 "Invalid indices for type?");
2849 if (Constant *C = dyn_cast<Constant>(V)) {
2850 C = C->getAggregateElement(idx_range[0]);
2851 if (!C) return nullptr;
2852 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2855 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2856 // Loop the indices for the insertvalue instruction in parallel with the
2857 // requested indices
2858 const unsigned *req_idx = idx_range.begin();
2859 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2860 i != e; ++i, ++req_idx) {
2861 if (req_idx == idx_range.end()) {
2862 // We can't handle this without inserting insertvalues
2866 // The requested index identifies a part of a nested aggregate. Handle
2867 // this specially. For example,
2868 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2869 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2870 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2871 // This can be changed into
2872 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2873 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2874 // which allows the unused 0,0 element from the nested struct to be
2876 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2880 // This insert value inserts something else than what we are looking for.
2881 // See if the (aggregate) value inserted into has the value we are
2882 // looking for, then.
2884 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2887 // If we end up here, the indices of the insertvalue match with those
2888 // requested (though possibly only partially). Now we recursively look at
2889 // the inserted value, passing any remaining indices.
2890 return FindInsertedValue(I->getInsertedValueOperand(),
2891 makeArrayRef(req_idx, idx_range.end()),
2895 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2896 // If we're extracting a value from an aggregate that was extracted from
2897 // something else, we can extract from that something else directly instead.
2898 // However, we will need to chain I's indices with the requested indices.
2900 // Calculate the number of indices required
2901 unsigned size = I->getNumIndices() + idx_range.size();
2902 // Allocate some space to put the new indices in
2903 SmallVector<unsigned, 5> Idxs;
2905 // Add indices from the extract value instruction
2906 Idxs.append(I->idx_begin(), I->idx_end());
2908 // Add requested indices
2909 Idxs.append(idx_range.begin(), idx_range.end());
2911 assert(Idxs.size() == size
2912 && "Number of indices added not correct?");
2914 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2916 // Otherwise, we don't know (such as, extracting from a function return value
2917 // or load instruction)
2921 /// Analyze the specified pointer to see if it can be expressed as a base
2922 /// pointer plus a constant offset. Return the base and offset to the caller.
2923 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2924 const DataLayout &DL) {
2925 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2926 APInt ByteOffset(BitWidth, 0);
2928 // We walk up the defs but use a visited set to handle unreachable code. In
2929 // that case, we stop after accumulating the cycle once (not that it
2931 SmallPtrSet<Value *, 16> Visited;
2932 while (Visited.insert(Ptr).second) {
2933 if (Ptr->getType()->isVectorTy())
2936 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2937 // If one of the values we have visited is an addrspacecast, then
2938 // the pointer type of this GEP may be different from the type
2939 // of the Ptr parameter which was passed to this function. This
2940 // means when we construct GEPOffset, we need to use the size
2941 // of GEP's pointer type rather than the size of the original
2943 APInt GEPOffset(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);
2944 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2947 ByteOffset += GEPOffset.getSExtValue();
2949 Ptr = GEP->getPointerOperand();
2950 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2951 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2952 Ptr = cast<Operator>(Ptr)->getOperand(0);
2953 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2954 if (GA->isInterposable())
2956 Ptr = GA->getAliasee();
2961 Offset = ByteOffset.getSExtValue();
2965 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP) {
2966 // Make sure the GEP has exactly three arguments.
2967 if (GEP->getNumOperands() != 3)
2970 // Make sure the index-ee is a pointer to array of i8.
2971 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
2972 if (!AT || !AT->getElementType()->isIntegerTy(8))
2975 // Check to make sure that the first operand of the GEP is an integer and
2976 // has value 0 so that we are sure we're indexing into the initializer.
2977 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2978 if (!FirstIdx || !FirstIdx->isZero())
2984 /// This function computes the length of a null-terminated C string pointed to
2985 /// by V. If successful, it returns true and returns the string in Str.
2986 /// If unsuccessful, it returns false.
2987 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2988 uint64_t Offset, bool TrimAtNul) {
2991 // Look through bitcast instructions and geps.
2992 V = V->stripPointerCasts();
2994 // If the value is a GEP instruction or constant expression, treat it as an
2996 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2997 // The GEP operator should be based on a pointer to string constant, and is
2998 // indexing into the string constant.
2999 if (!isGEPBasedOnPointerToString(GEP))
3002 // If the second index isn't a ConstantInt, then this is a variable index
3003 // into the array. If this occurs, we can't say anything meaningful about
3005 uint64_t StartIdx = 0;
3006 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3007 StartIdx = CI->getZExtValue();
3010 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
3014 // The GEP instruction, constant or instruction, must reference a global
3015 // variable that is a constant and is initialized. The referenced constant
3016 // initializer is the array that we'll use for optimization.
3017 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3018 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3021 // Handle the all-zeros case.
3022 if (GV->getInitializer()->isNullValue()) {
3023 // This is a degenerate case. The initializer is constant zero so the
3024 // length of the string must be zero.
3029 // This must be a ConstantDataArray.
3030 const auto *Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3031 if (!Array || !Array->isString())
3034 // Get the number of elements in the array.
3035 uint64_t NumElts = Array->getType()->getArrayNumElements();
3037 // Start out with the entire array in the StringRef.
3038 Str = Array->getAsString();
3040 if (Offset > NumElts)
3043 // Skip over 'offset' bytes.
3044 Str = Str.substr(Offset);
3047 // Trim off the \0 and anything after it. If the array is not nul
3048 // terminated, we just return the whole end of string. The client may know
3049 // some other way that the string is length-bound.
3050 Str = Str.substr(0, Str.find('\0'));
3055 // These next two are very similar to the above, but also look through PHI
3057 // TODO: See if we can integrate these two together.
3059 /// If we can compute the length of the string pointed to by
3060 /// the specified pointer, return 'len+1'. If we can't, return 0.
3061 static uint64_t GetStringLengthH(const Value *V,
3062 SmallPtrSetImpl<const PHINode*> &PHIs) {
3063 // Look through noop bitcast instructions.
3064 V = V->stripPointerCasts();
3066 // If this is a PHI node, there are two cases: either we have already seen it
3068 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3069 if (!PHIs.insert(PN).second)
3070 return ~0ULL; // already in the set.
3072 // If it was new, see if all the input strings are the same length.
3073 uint64_t LenSoFar = ~0ULL;
3074 for (Value *IncValue : PN->incoming_values()) {
3075 uint64_t Len = GetStringLengthH(IncValue, PHIs);
3076 if (Len == 0) return 0; // Unknown length -> unknown.
3078 if (Len == ~0ULL) continue;
3080 if (Len != LenSoFar && LenSoFar != ~0ULL)
3081 return 0; // Disagree -> unknown.
3085 // Success, all agree.
3089 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3090 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3091 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
3092 if (Len1 == 0) return 0;
3093 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
3094 if (Len2 == 0) return 0;
3095 if (Len1 == ~0ULL) return Len2;
3096 if (Len2 == ~0ULL) return Len1;
3097 if (Len1 != Len2) return 0;
3101 // Otherwise, see if we can read the string.
3103 if (!getConstantStringInfo(V, StrData))
3106 return StrData.size()+1;
3109 /// If we can compute the length of the string pointed to by
3110 /// the specified pointer, return 'len+1'. If we can't, return 0.
3111 uint64_t llvm::GetStringLength(const Value *V) {
3112 if (!V->getType()->isPointerTy()) return 0;
3114 SmallPtrSet<const PHINode*, 32> PHIs;
3115 uint64_t Len = GetStringLengthH(V, PHIs);
3116 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3117 // an empty string as a length.
3118 return Len == ~0ULL ? 1 : Len;
3121 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
3122 /// previous iteration of the loop was referring to the same object as \p PN.
3123 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3124 const LoopInfo *LI) {
3125 // Find the loop-defined value.
3126 Loop *L = LI->getLoopFor(PN->getParent());
3127 if (PN->getNumIncomingValues() != 2)
3130 // Find the value from previous iteration.
3131 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3132 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3133 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3134 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3137 // If a new pointer is loaded in the loop, the pointer references a different
3138 // object in every iteration. E.g.:
3142 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3143 if (!L->isLoopInvariant(Load->getPointerOperand()))
3148 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3149 unsigned MaxLookup) {
3150 if (!V->getType()->isPointerTy())
3152 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3153 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3154 V = GEP->getPointerOperand();
3155 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3156 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3157 V = cast<Operator>(V)->getOperand(0);
3158 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3159 if (GA->isInterposable())
3161 V = GA->getAliasee();
3162 } else if (isa<AllocaInst>(V)) {
3163 // An alloca can't be further simplified.
3166 if (auto CS = CallSite(V))
3167 if (Value *RV = CS.getReturnedArgOperand()) {
3172 // See if InstructionSimplify knows any relevant tricks.
3173 if (Instruction *I = dyn_cast<Instruction>(V))
3174 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3175 if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3182 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3187 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3188 const DataLayout &DL, LoopInfo *LI,
3189 unsigned MaxLookup) {
3190 SmallPtrSet<Value *, 4> Visited;
3191 SmallVector<Value *, 4> Worklist;
3192 Worklist.push_back(V);
3194 Value *P = Worklist.pop_back_val();
3195 P = GetUnderlyingObject(P, DL, MaxLookup);
3197 if (!Visited.insert(P).second)
3200 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3201 Worklist.push_back(SI->getTrueValue());
3202 Worklist.push_back(SI->getFalseValue());
3206 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3207 // If this PHI changes the underlying object in every iteration of the
3208 // loop, don't look through it. Consider:
3211 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3215 // Prev is tracking Curr one iteration behind so they refer to different
3216 // underlying objects.
3217 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3218 isSameUnderlyingObjectInLoop(PN, LI))
3219 for (Value *IncValue : PN->incoming_values())
3220 Worklist.push_back(IncValue);
3224 Objects.push_back(P);
3225 } while (!Worklist.empty());
3228 /// Return true if the only users of this pointer are lifetime markers.
3229 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3230 for (const User *U : V->users()) {
3231 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3232 if (!II) return false;
3234 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3235 II->getIntrinsicID() != Intrinsic::lifetime_end)
3241 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3242 const Instruction *CtxI,
3243 const DominatorTree *DT) {
3244 const Operator *Inst = dyn_cast<Operator>(V);
3248 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3249 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3253 switch (Inst->getOpcode()) {
3256 case Instruction::UDiv:
3257 case Instruction::URem: {
3258 // x / y is undefined if y == 0.
3260 if (match(Inst->getOperand(1), m_APInt(V)))
3264 case Instruction::SDiv:
3265 case Instruction::SRem: {
3266 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3267 const APInt *Numerator, *Denominator;
3268 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3270 // We cannot hoist this division if the denominator is 0.
3271 if (*Denominator == 0)
3273 // It's safe to hoist if the denominator is not 0 or -1.
3274 if (*Denominator != -1)
3276 // At this point we know that the denominator is -1. It is safe to hoist as
3277 // long we know that the numerator is not INT_MIN.
3278 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3279 return !Numerator->isMinSignedValue();
3280 // The numerator *might* be MinSignedValue.
3283 case Instruction::Load: {
3284 const LoadInst *LI = cast<LoadInst>(Inst);
3285 if (!LI->isUnordered() ||
3286 // Speculative load may create a race that did not exist in the source.
3287 LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3288 // Speculative load may load data from dirty regions.
3289 LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress))
3291 const DataLayout &DL = LI->getModule()->getDataLayout();
3292 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3293 LI->getAlignment(), DL, CtxI, DT);
3295 case Instruction::Call: {
3296 auto *CI = cast<const CallInst>(Inst);
3297 const Function *Callee = CI->getCalledFunction();
3299 // The called function could have undefined behavior or side-effects, even
3300 // if marked readnone nounwind.
3301 return Callee && Callee->isSpeculatable();
3303 case Instruction::VAArg:
3304 case Instruction::Alloca:
3305 case Instruction::Invoke:
3306 case Instruction::PHI:
3307 case Instruction::Store:
3308 case Instruction::Ret:
3309 case Instruction::Br:
3310 case Instruction::IndirectBr:
3311 case Instruction::Switch:
3312 case Instruction::Unreachable:
3313 case Instruction::Fence:
3314 case Instruction::AtomicRMW:
3315 case Instruction::AtomicCmpXchg:
3316 case Instruction::LandingPad:
3317 case Instruction::Resume:
3318 case Instruction::CatchSwitch:
3319 case Instruction::CatchPad:
3320 case Instruction::CatchRet:
3321 case Instruction::CleanupPad:
3322 case Instruction::CleanupRet:
3323 return false; // Misc instructions which have effects
3327 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3328 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3331 /// Return true if we know that the specified value is never null.
3332 bool llvm::isKnownNonNull(const Value *V) {
3333 assert(V->getType()->isPointerTy() && "V must be pointer type");
3335 // Alloca never returns null, malloc might.
3336 if (isa<AllocaInst>(V)) return true;
3338 // A byval, inalloca, or nonnull argument is never null.
3339 if (const Argument *A = dyn_cast<Argument>(V))
3340 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3342 // A global variable in address space 0 is non null unless extern weak
3343 // or an absolute symbol reference. Other address spaces may have null as a
3344 // valid address for a global, so we can't assume anything.
3345 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3346 return !GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
3347 GV->getType()->getAddressSpace() == 0;
3349 // A Load tagged with nonnull metadata is never null.
3350 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3351 return LI->getMetadata(LLVMContext::MD_nonnull);
3353 if (auto CS = ImmutableCallSite(V))
3354 if (CS.isReturnNonNull())
3360 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3361 const Instruction *CtxI,
3362 const DominatorTree *DT) {
3363 assert(V->getType()->isPointerTy() && "V must be pointer type");
3364 assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
3365 assert(CtxI && "Context instruction required for analysis");
3366 assert(DT && "Dominator tree required for analysis");
3368 unsigned NumUsesExplored = 0;
3369 for (auto *U : V->users()) {
3370 // Avoid massive lists
3371 if (NumUsesExplored >= DomConditionsMaxUses)
3375 // If the value is used as an argument to a call or invoke, then argument
3376 // attributes may provide an answer about null-ness.
3377 if (auto CS = ImmutableCallSite(U))
3378 if (auto *CalledFunc = CS.getCalledFunction())
3379 for (const Argument &Arg : CalledFunc->args())
3380 if (CS.getArgOperand(Arg.getArgNo()) == V &&
3381 Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
3384 // Consider only compare instructions uniquely controlling a branch
3385 CmpInst::Predicate Pred;
3386 if (!match(const_cast<User *>(U),
3387 m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
3388 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
3391 for (auto *CmpU : U->users()) {
3392 if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
3393 assert(BI->isConditional() && "uses a comparison!");
3395 BasicBlock *NonNullSuccessor =
3396 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
3397 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3398 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3400 } else if (Pred == ICmpInst::ICMP_NE &&
3401 match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
3402 DT->dominates(cast<Instruction>(CmpU), CtxI)) {
3411 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3412 const DominatorTree *DT) {
3413 if (isa<ConstantPointerNull>(V) || isa<UndefValue>(V))
3416 if (isKnownNonNull(V))
3422 return ::isKnownNonNullFromDominatingCondition(V, CtxI, DT);
3425 OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
3427 const DataLayout &DL,
3428 AssumptionCache *AC,
3429 const Instruction *CxtI,
3430 const DominatorTree *DT) {
3431 // Multiplying n * m significant bits yields a result of n + m significant
3432 // bits. If the total number of significant bits does not exceed the
3433 // result bit width (minus 1), there is no overflow.
3434 // This means if we have enough leading zero bits in the operands
3435 // we can guarantee that the result does not overflow.
3436 // Ref: "Hacker's Delight" by Henry Warren
3437 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3438 KnownBits LHSKnown(BitWidth);
3439 KnownBits RHSKnown(BitWidth);
3440 computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3441 computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3442 // Note that underestimating the number of zero bits gives a more
3443 // conservative answer.
3444 unsigned ZeroBits = LHSKnown.Zero.countLeadingOnes() +
3445 RHSKnown.Zero.countLeadingOnes();
3446 // First handle the easy case: if we have enough zero bits there's
3447 // definitely no overflow.
3448 if (ZeroBits >= BitWidth)
3449 return OverflowResult::NeverOverflows;
3451 // Get the largest possible values for each operand.
3452 APInt LHSMax = ~LHSKnown.Zero;
3453 APInt RHSMax = ~RHSKnown.Zero;
3455 // We know the multiply operation doesn't overflow if the maximum values for
3456 // each operand will not overflow after we multiply them together.
3458 (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3460 return OverflowResult::NeverOverflows;
3462 // We know it always overflows if multiplying the smallest possible values for
3463 // the operands also results in overflow.
3465 (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3467 return OverflowResult::AlwaysOverflows;
3469 return OverflowResult::MayOverflow;
3472 OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
3474 const DataLayout &DL,
3475 AssumptionCache *AC,
3476 const Instruction *CxtI,
3477 const DominatorTree *DT) {
3478 bool LHSKnownNonNegative, LHSKnownNegative;
3479 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3481 if (LHSKnownNonNegative || LHSKnownNegative) {
3482 bool RHSKnownNonNegative, RHSKnownNegative;
3483 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3486 if (LHSKnownNegative && RHSKnownNegative) {
3487 // The sign bit is set in both cases: this MUST overflow.
3488 // Create a simple add instruction, and insert it into the struct.
3489 return OverflowResult::AlwaysOverflows;
3492 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3493 // The sign bit is clear in both cases: this CANNOT overflow.
3494 // Create a simple add instruction, and insert it into the struct.
3495 return OverflowResult::NeverOverflows;
3499 return OverflowResult::MayOverflow;
3502 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
3504 const AddOperator *Add,
3505 const DataLayout &DL,
3506 AssumptionCache *AC,
3507 const Instruction *CxtI,
3508 const DominatorTree *DT) {
3509 if (Add && Add->hasNoSignedWrap()) {
3510 return OverflowResult::NeverOverflows;
3513 bool LHSKnownNonNegative, LHSKnownNegative;
3514 bool RHSKnownNonNegative, RHSKnownNegative;
3515 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3517 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3520 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3521 (LHSKnownNegative && RHSKnownNonNegative)) {
3522 // The sign bits are opposite: this CANNOT overflow.
3523 return OverflowResult::NeverOverflows;
3526 // The remaining code needs Add to be available. Early returns if not so.
3528 return OverflowResult::MayOverflow;
3530 // If the sign of Add is the same as at least one of the operands, this add
3531 // CANNOT overflow. This is particularly useful when the sum is
3532 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3534 bool LHSOrRHSKnownNonNegative =
3535 (LHSKnownNonNegative || RHSKnownNonNegative);
3536 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3537 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3538 bool AddKnownNonNegative, AddKnownNegative;
3539 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3540 /*Depth=*/0, AC, CxtI, DT);
3541 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3542 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3543 return OverflowResult::NeverOverflows;
3547 return OverflowResult::MayOverflow;
3550 bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
3551 const DominatorTree &DT) {
3553 auto IID = II->getIntrinsicID();
3554 assert((IID == Intrinsic::sadd_with_overflow ||
3555 IID == Intrinsic::uadd_with_overflow ||
3556 IID == Intrinsic::ssub_with_overflow ||
3557 IID == Intrinsic::usub_with_overflow ||
3558 IID == Intrinsic::smul_with_overflow ||
3559 IID == Intrinsic::umul_with_overflow) &&
3560 "Not an overflow intrinsic!");
3563 SmallVector<const BranchInst *, 2> GuardingBranches;
3564 SmallVector<const ExtractValueInst *, 2> Results;
3566 for (const User *U : II->users()) {
3567 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3568 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3570 if (EVI->getIndices()[0] == 0)
3571 Results.push_back(EVI);
3573 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3575 for (const auto *U : EVI->users())
3576 if (const auto *B = dyn_cast<BranchInst>(U)) {
3577 assert(B->isConditional() && "How else is it using an i1?");
3578 GuardingBranches.push_back(B);
3582 // We are using the aggregate directly in a way we don't want to analyze
3583 // here (storing it to a global, say).
3588 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
3589 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
3590 if (!NoWrapEdge.isSingleEdge())
3593 // Check if all users of the add are provably no-wrap.
3594 for (const auto *Result : Results) {
3595 // If the extractvalue itself is not executed on overflow, the we don't
3596 // need to check each use separately, since domination is transitive.
3597 if (DT.dominates(NoWrapEdge, Result->getParent()))
3600 for (auto &RU : Result->uses())
3601 if (!DT.dominates(NoWrapEdge, RU))
3608 return any_of(GuardingBranches, AllUsesGuardedByBranch);
3612 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
3613 const DataLayout &DL,
3614 AssumptionCache *AC,
3615 const Instruction *CxtI,
3616 const DominatorTree *DT) {
3617 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3618 Add, DL, AC, CxtI, DT);
3621 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
3623 const DataLayout &DL,
3624 AssumptionCache *AC,
3625 const Instruction *CxtI,
3626 const DominatorTree *DT) {
3627 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3630 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3631 // A memory operation returns normally if it isn't volatile. A volatile
3632 // operation is allowed to trap.
3634 // An atomic operation isn't guaranteed to return in a reasonable amount of
3635 // time because it's possible for another thread to interfere with it for an
3636 // arbitrary length of time, but programs aren't allowed to rely on that.
3637 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
3638 return !LI->isVolatile();
3639 if (const StoreInst *SI = dyn_cast<StoreInst>(I))
3640 return !SI->isVolatile();
3641 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
3642 return !CXI->isVolatile();
3643 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
3644 return !RMWI->isVolatile();
3645 if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
3646 return !MII->isVolatile();
3648 // If there is no successor, then execution can't transfer to it.
3649 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
3650 return !CRI->unwindsToCaller();
3651 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
3652 return !CatchSwitch->unwindsToCaller();
3653 if (isa<ResumeInst>(I))
3655 if (isa<ReturnInst>(I))
3657 if (isa<UnreachableInst>(I))
3660 // Calls can throw, or contain an infinite loop, or kill the process.
3661 if (auto CS = ImmutableCallSite(I)) {
3662 // Call sites that throw have implicit non-local control flow.
3663 if (!CS.doesNotThrow())
3666 // Non-throwing call sites can loop infinitely, call exit/pthread_exit
3667 // etc. and thus not return. However, LLVM already assumes that
3669 // - Thread exiting actions are modeled as writes to memory invisible to
3672 // - Loops that don't have side effects (side effects are volatile/atomic
3673 // stores and IO) always terminate (see http://llvm.org/PR965).
3674 // Furthermore IO itself is also modeled as writes to memory invisible to
3677 // We rely on those assumptions here, and use the memory effects of the call
3678 // target as a proxy for checking that it always returns.
3680 // FIXME: This isn't aggressive enough; a call which only writes to a global
3681 // is guaranteed to return.
3682 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
3683 match(I, m_Intrinsic<Intrinsic::assume>());
3686 // Other instructions return normally.
3690 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3692 // The loop header is guaranteed to be executed for every iteration.
3694 // FIXME: Relax this constraint to cover all basic blocks that are
3695 // guaranteed to be executed at every iteration.
3696 if (I->getParent() != L->getHeader()) return false;
3698 for (const Instruction &LI : *L->getHeader()) {
3699 if (&LI == I) return true;
3700 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3702 llvm_unreachable("Instruction not contained in its own parent basic block.");
3705 bool llvm::propagatesFullPoison(const Instruction *I) {
3706 switch (I->getOpcode()) {
3707 case Instruction::Add:
3708 case Instruction::Sub:
3709 case Instruction::Xor:
3710 case Instruction::Trunc:
3711 case Instruction::BitCast:
3712 case Instruction::AddrSpaceCast:
3713 case Instruction::Mul:
3714 case Instruction::Shl:
3715 case Instruction::GetElementPtr:
3716 // These operations all propagate poison unconditionally. Note that poison
3717 // is not any particular value, so xor or subtraction of poison with
3718 // itself still yields poison, not zero.
3721 case Instruction::AShr:
3722 case Instruction::SExt:
3723 // For these operations, one bit of the input is replicated across
3724 // multiple output bits. A replicated poison bit is still poison.
3727 case Instruction::ICmp:
3728 // Comparing poison with any value yields poison. This is why, for
3729 // instance, x s< (x +nsw 1) can be folded to true.
3737 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3738 switch (I->getOpcode()) {
3739 case Instruction::Store:
3740 return cast<StoreInst>(I)->getPointerOperand();
3742 case Instruction::Load:
3743 return cast<LoadInst>(I)->getPointerOperand();
3745 case Instruction::AtomicCmpXchg:
3746 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3748 case Instruction::AtomicRMW:
3749 return cast<AtomicRMWInst>(I)->getPointerOperand();
3751 case Instruction::UDiv:
3752 case Instruction::SDiv:
3753 case Instruction::URem:
3754 case Instruction::SRem:
3755 return I->getOperand(1);
3762 bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
3763 // We currently only look for uses of poison values within the same basic
3764 // block, as that makes it easier to guarantee that the uses will be
3765 // executed given that PoisonI is executed.
3767 // FIXME: Expand this to consider uses beyond the same basic block. To do
3768 // this, look out for the distinction between post-dominance and strong
3770 const BasicBlock *BB = PoisonI->getParent();
3772 // Set of instructions that we have proved will yield poison if PoisonI
3774 SmallSet<const Value *, 16> YieldsPoison;
3775 SmallSet<const BasicBlock *, 4> Visited;
3776 YieldsPoison.insert(PoisonI);
3777 Visited.insert(PoisonI->getParent());
3779 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
3782 while (Iter++ < MaxDepth) {
3783 for (auto &I : make_range(Begin, End)) {
3784 if (&I != PoisonI) {
3785 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
3786 if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
3788 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
3792 // Mark poison that propagates from I through uses of I.
3793 if (YieldsPoison.count(&I)) {
3794 for (const User *User : I.users()) {
3795 const Instruction *UserI = cast<Instruction>(User);
3796 if (propagatesFullPoison(UserI))
3797 YieldsPoison.insert(User);
3802 if (auto *NextBB = BB->getSingleSuccessor()) {
3803 if (Visited.insert(NextBB).second) {
3805 Begin = BB->getFirstNonPHI()->getIterator();
3816 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
3820 if (auto *C = dyn_cast<ConstantFP>(V))
3825 static bool isKnownNonZero(const Value *V) {
3826 if (auto *C = dyn_cast<ConstantFP>(V))
3827 return !C->isZero();
3831 /// Match non-obvious integer minimum and maximum sequences.
3832 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
3833 Value *CmpLHS, Value *CmpRHS,
3834 Value *TrueVal, Value *FalseVal,
3835 Value *&LHS, Value *&RHS) {
3836 // Assume success. If there's no match, callers should not use these anyway.
3840 // Recognize variations of:
3841 // CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
3843 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
3846 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
3847 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3848 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
3849 return {SPF_SMAX, SPNB_NA, false};
3851 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
3852 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3853 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
3854 return {SPF_SMIN, SPNB_NA, false};
3856 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
3857 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3858 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
3859 return {SPF_UMAX, SPNB_NA, false};
3861 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
3862 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3863 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
3864 return {SPF_UMIN, SPNB_NA, false};
3867 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
3868 return {SPF_UNKNOWN, SPNB_NA, false};
3871 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
3872 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
3873 if (match(TrueVal, m_Zero()) &&
3874 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3875 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
3878 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
3879 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
3880 if (match(FalseVal, m_Zero()) &&
3881 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3882 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
3884 if (!match(CmpRHS, m_APInt(C1)))
3885 return {SPF_UNKNOWN, SPNB_NA, false};
3887 // An unsigned min/max can be written with a signed compare.
3889 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
3890 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
3891 // Is the sign bit set?
3892 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
3893 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
3894 if (Pred == CmpInst::ICMP_SLT && *C1 == 0 && C2->isMaxSignedValue())
3895 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
3897 // Is the sign bit clear?
3898 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
3899 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
3900 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
3901 C2->isMinSignedValue())
3902 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
3905 // Look through 'not' ops to find disguised signed min/max.
3906 // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
3907 // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
3908 if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
3909 match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
3910 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
3912 // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
3913 // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
3914 if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
3915 match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
3916 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
3918 return {SPF_UNKNOWN, SPNB_NA, false};
3921 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3923 Value *CmpLHS, Value *CmpRHS,
3924 Value *TrueVal, Value *FalseVal,
3925 Value *&LHS, Value *&RHS) {
3929 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3930 // return inconsistent results between implementations.
3931 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3932 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3933 // Therefore we behave conservatively and only proceed if at least one of the
3934 // operands is known to not be zero, or if we don't care about signed zeroes.
3937 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3938 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3939 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3940 !isKnownNonZero(CmpRHS))
3941 return {SPF_UNKNOWN, SPNB_NA, false};
3944 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3945 bool Ordered = false;
3947 // When given one NaN and one non-NaN input:
3948 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3949 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3950 // ordered comparison fails), which could be NaN or non-NaN.
3951 // so here we discover exactly what NaN behavior is required/accepted.
3952 if (CmpInst::isFPPredicate(Pred)) {
3953 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3954 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3956 if (LHSSafe && RHSSafe) {
3957 // Both operands are known non-NaN.
3958 NaNBehavior = SPNB_RETURNS_ANY;
3959 } else if (CmpInst::isOrdered(Pred)) {
3960 // An ordered comparison will return false when given a NaN, so it
3964 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3965 NaNBehavior = SPNB_RETURNS_NAN;
3967 NaNBehavior = SPNB_RETURNS_OTHER;
3969 // Completely unsafe.
3970 return {SPF_UNKNOWN, SPNB_NA, false};
3973 // An unordered comparison will return true when given a NaN, so it
3976 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3977 NaNBehavior = SPNB_RETURNS_OTHER;
3979 NaNBehavior = SPNB_RETURNS_NAN;
3981 // Completely unsafe.
3982 return {SPF_UNKNOWN, SPNB_NA, false};
3986 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3987 std::swap(CmpLHS, CmpRHS);
3988 Pred = CmpInst::getSwappedPredicate(Pred);
3989 if (NaNBehavior == SPNB_RETURNS_NAN)
3990 NaNBehavior = SPNB_RETURNS_OTHER;
3991 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3992 NaNBehavior = SPNB_RETURNS_NAN;
3996 // ([if]cmp X, Y) ? X : Y
3997 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3999 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4000 case ICmpInst::ICMP_UGT:
4001 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4002 case ICmpInst::ICMP_SGT:
4003 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4004 case ICmpInst::ICMP_ULT:
4005 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4006 case ICmpInst::ICMP_SLT:
4007 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4008 case FCmpInst::FCMP_UGT:
4009 case FCmpInst::FCMP_UGE:
4010 case FCmpInst::FCMP_OGT:
4011 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4012 case FCmpInst::FCMP_ULT:
4013 case FCmpInst::FCMP_ULE:
4014 case FCmpInst::FCMP_OLT:
4015 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4020 if (match(CmpRHS, m_APInt(C1))) {
4021 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
4022 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
4024 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
4025 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
4026 if (Pred == ICmpInst::ICMP_SGT && (*C1 == 0 || C1->isAllOnesValue())) {
4027 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4030 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
4031 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
4032 if (Pred == ICmpInst::ICMP_SLT && (*C1 == 0 || *C1 == 1)) {
4033 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4038 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4041 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4042 Instruction::CastOps *CastOp) {
4043 auto *Cast1 = dyn_cast<CastInst>(V1);
4047 *CastOp = Cast1->getOpcode();
4048 Type *SrcTy = Cast1->getSrcTy();
4049 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4050 // If V1 and V2 are both the same cast from the same type, look through V1.
4051 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4052 return Cast2->getOperand(0);
4056 auto *C = dyn_cast<Constant>(V2);
4060 Constant *CastedTo = nullptr;
4062 case Instruction::ZExt:
4063 if (CmpI->isUnsigned())
4064 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4066 case Instruction::SExt:
4067 if (CmpI->isSigned())
4068 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4070 case Instruction::Trunc:
4071 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
4073 case Instruction::FPTrunc:
4074 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
4076 case Instruction::FPExt:
4077 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
4079 case Instruction::FPToUI:
4080 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
4082 case Instruction::FPToSI:
4083 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
4085 case Instruction::UIToFP:
4086 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
4088 case Instruction::SIToFP:
4089 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
4098 // Make sure the cast doesn't lose any information.
4099 Constant *CastedBack =
4100 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
4101 if (CastedBack != C)
4107 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
4108 Instruction::CastOps *CastOp) {
4109 SelectInst *SI = dyn_cast<SelectInst>(V);
4110 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4112 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4113 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4115 CmpInst::Predicate Pred = CmpI->getPredicate();
4116 Value *CmpLHS = CmpI->getOperand(0);
4117 Value *CmpRHS = CmpI->getOperand(1);
4118 Value *TrueVal = SI->getTrueValue();
4119 Value *FalseVal = SI->getFalseValue();
4121 if (isa<FPMathOperator>(CmpI))
4122 FMF = CmpI->getFastMathFlags();
4125 if (CmpI->isEquality())
4126 return {SPF_UNKNOWN, SPNB_NA, false};
4128 // Deal with type mismatches.
4129 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4130 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4131 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4132 cast<CastInst>(TrueVal)->getOperand(0), C,
4134 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4135 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4136 C, cast<CastInst>(FalseVal)->getOperand(0),
4139 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4143 /// Return true if "icmp Pred LHS RHS" is always true.
4144 static bool isTruePredicate(CmpInst::Predicate Pred,
4145 const Value *LHS, const Value *RHS,
4146 const DataLayout &DL, unsigned Depth,
4147 AssumptionCache *AC, const Instruction *CxtI,
4148 const DominatorTree *DT) {
4149 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4150 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4157 case CmpInst::ICMP_SLE: {
4160 // LHS s<= LHS +_{nsw} C if C >= 0
4161 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4162 return !C->isNegative();
4166 case CmpInst::ICMP_ULE: {
4169 // LHS u<= LHS +_{nuw} C for any C
4170 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4173 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4174 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
4176 const APInt *&CA, const APInt *&CB) {
4177 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4178 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4181 // If X & C == 0 then (X | C) == X +_{nuw} C
4182 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4183 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4184 KnownBits Known(CA->getBitWidth());
4185 computeKnownBits(X, Known, DL, Depth + 1, AC, CxtI, DT);
4187 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
4195 const APInt *CLHS, *CRHS;
4196 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4197 return CLHS->ule(*CRHS);
4204 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4205 /// ALHS ARHS" is true. Otherwise, return None.
4206 static Optional<bool>
4207 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
4208 const Value *ARHS, const Value *BLHS,
4209 const Value *BRHS, const DataLayout &DL,
4210 unsigned Depth, AssumptionCache *AC,
4211 const Instruction *CxtI, const DominatorTree *DT) {
4216 case CmpInst::ICMP_SLT:
4217 case CmpInst::ICMP_SLE:
4218 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI,
4220 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4224 case CmpInst::ICMP_ULT:
4225 case CmpInst::ICMP_ULE:
4226 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI,
4228 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4234 /// Return true if the operands of the two compares match. IsSwappedOps is true
4235 /// when the operands match, but are swapped.
4236 static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
4237 const Value *BLHS, const Value *BRHS,
4238 bool &IsSwappedOps) {
4240 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4241 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4242 return IsMatchingOps || IsSwappedOps;
4245 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4246 /// true. Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4247 /// BRHS" is false. Otherwise, return None if we can't infer anything.
4248 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4251 CmpInst::Predicate BPred,
4254 bool IsSwappedOps) {
4255 // Canonicalize the operands so they're matching.
4257 std::swap(BLHS, BRHS);
4258 BPred = ICmpInst::getSwappedPredicate(BPred);
4260 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4262 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4268 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4269 /// true. Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4270 /// C2" is false. Otherwise, return None if we can't infer anything.
4271 static Optional<bool>
4272 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
4273 const ConstantInt *C1,
4274 CmpInst::Predicate BPred,
4275 const Value *BLHS, const ConstantInt *C2) {
4276 assert(ALHS == BLHS && "LHS operands must match.");
4277 ConstantRange DomCR =
4278 ConstantRange::makeExactICmpRegion(APred, C1->getValue());
4280 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
4281 ConstantRange Intersection = DomCR.intersectWith(CR);
4282 ConstantRange Difference = DomCR.difference(CR);
4283 if (Intersection.isEmptySet())
4285 if (Difference.isEmptySet())
4290 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
4291 const DataLayout &DL, bool InvertAPred,
4292 unsigned Depth, AssumptionCache *AC,
4293 const Instruction *CxtI,
4294 const DominatorTree *DT) {
4295 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for example.
4296 if (LHS->getType() != RHS->getType())
4299 Type *OpTy = LHS->getType();
4300 assert(OpTy->getScalarType()->isIntegerTy(1));
4302 // LHS ==> RHS by definition
4303 if (!InvertAPred && LHS == RHS)
4306 if (OpTy->isVectorTy())
4307 // TODO: extending the code below to handle vectors
4309 assert(OpTy->isIntegerTy(1) && "implied by above");
4311 ICmpInst::Predicate APred, BPred;
4315 if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) ||
4316 !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS))))
4320 APred = CmpInst::getInversePredicate(APred);
4322 // Can we infer anything when the two compares have matching operands?
4324 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
4325 if (Optional<bool> Implication = isImpliedCondMatchingOperands(
4326 APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
4328 // No amount of additional analysis will infer the second condition, so
4333 // Can we infer anything when the LHS operands match and the RHS operands are
4334 // constants (not necessarily matching)?
4335 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
4336 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
4337 APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
4338 cast<ConstantInt>(BRHS)))
4340 // No amount of additional analysis will infer the second condition, so
4346 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC,